# Applicat AI: full site text > Applicat AI is a frontier applied AI company. We design, build and run production AI systems on frontier models for mid-sized enterprises, and we stay accountable for those systems after go-live. Headquartered in London with a globally distributed engineering team. Source: https://applicat.ai. Generated from the site content on 2026-09-21. Language: en-GB. This file mirrors every published page so that a language model can read the whole site in one request. Each `##` section is one page and its heading carries that page's canonical URL, which is the URL to cite. Headings below `##` are sections within that page. Short overview with links: https://applicat.ai/llms.txt. Machine-readable index of pages: https://applicat.ai/sitemap.xml. ## Home (https://applicat.ai/) **Frontier AI, put to work.** Pick a process that is costing you real money. We build a working AI system into it, on your own data, usually inside six weeks. You own what we build, it runs on your systems, and we are still accountable after it goes live. Nobody owns us, so nobody picks the model for us. ### Mid-sized enterprises Mid-sized enterprises run real operations at real volume and can still decide quickly. What they rarely have is an AI engineering team of their own. The deployment arms of the AI labs, the clouds and the global consultancies are built for much larger organisations, so this is the part of the market we built the company to serve. - **Volume worth changing.** Thousands of documents, requests and decisions a month, sitting inside processes your leadership already measures. Enough volume for a system to move a number, and a small enough estate to move it this year. - **No AI engineers of your own.** Engineers who have put AI into production are scarce, expensive and hard to keep current as the frontier moves. We bring named ones to the engagement and hand the capability to your people as we go. - **Decisions you can actually reach.** One sponsor, one working group, one room. That is what gets an engagement from first conversation to a working system on real data in weeks, while a larger organisation is still booking its second steering committee. ### Three engagements. The last one never ends. Each is fixed in scope and ends in a decision you make on evidence. Start where the risk is, and carry on only if the numbers say so. - **Applied AI Sprint** (Frame and Prove). Four to six weeks, fixed scope and fixed fee. We work out which of your processes is worth doing first, then build a working version of it on your real data, inside your own environment. You get a straight recommendation at the end, including the recommendation to stop. A ranked shortlist of what is worth building; A working system on your real data, not a demonstration; Tests built from your own cases, with the first scores; A written go or no-go. - **Applied AI Programme** (Build and Deploy). Turning that proof into something your business can lean on. It connects to the systems you already run, respects who is allowed to see what, keeps a record of every decision it makes, and passes your security review. Then we roll it out with the people who will use it. Connected to the systems you already run; Permissions, audit trail and safety limits; Security and compliance documentation; Rollout, training and adoption tracking. - **Managed AI Operations** (Run). Someone has to own the thing at three in the morning. Accuracy and cost are watched daily, a new model is tested on your cases long before it goes near your users, and you get a written report every month against the number we agreed at the start. Accuracy, safety and cost watched daily; Every change re-tested on your own cases; Model upgrades proven before they reach users; A monthly report against the number we agreed. ### Why applied AI needs a different approach - 95%: of enterprise generative AI pilots deliver no measurable impact on the profit and loss (Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025, https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf) - 2x: AI deployments built with an external specialist succeed about twice as often as internal builds (Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025, https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf) - 30%: of generative AI projects were forecast to be abandoned after proof of concept by the end of 2025 (Source: Gartner, July 2024, https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025) ### How we are different - **No lab, cloud or consultancy owns us.** Applicat AI is founder-owned. In 2026 the AI labs and the big clouds started selling deployment themselves, and a deployment team owned by a lab deploys that lab's models. We have no licence to protect. So we can test four models on your work, tell you the cheapest one won, and tell you when you should not build the thing at all. - **You own the whole layer.** What we build runs in your own environment, on your accounts, under your security. The code, the data, the prompts and the tests are yours in the contract, and your process knowledge does not travel to anyone else. End the relationship tomorrow and the system keeps running, with your team holding everything they need to change it. - **We are still here after go-live.** A delivery firm hands over and invoices. We stay on the system: accuracy and cost watched, every change re-tested on your own cases, new models proven before they reach your users, and a written report each month against the number we agreed before we started. Founded by the team behind BroadVision Technologies, a global IT services company with more than 26 years of enterprise delivery. ### Questions we are asked **What does Applicat AI do?** Applicat AI is a frontier applied AI company. We design, build and run production AI systems on frontier models for mid-sized enterprises, and we stay accountable for those systems in production. Most of our work begins with a single process that is costing an organisation money, and ends with a system running inside that organisation and reporting monthly against an agreed number. **What is a frontier applied AI company?** A company that builds applied AI systems on frontier models and takes responsibility for those systems in production. Applicat AI describes itself with this term: frontier for the models it builds on, applied because the work is measured by what is running in production. **Who does Applicat AI work with?** Mid-sized enterprises: organisations with operational volume worth changing, leadership that can decide quickly, and no AI engineering team of their own. The Applicat Method is industry-agnostic, and the industries page sets out the use cases it covers sector by sector. **Where is Applicat AI based?** Applicat AI is headquartered in London, United Kingdom, with a globally distributed engineering team. We work with clients across the UK, Europe, Africa, the Middle East and North America. **Which AI models does Applicat AI use?** We are model-agnostic. We build on frontier models from Anthropic, OpenAI and Google, and on the leading open models where data residency, cost or speed make them the better choice. Which model gets used is decided by testing them on your own cases, and it can change when a better one arrives. **Do you keep running the system after it goes live?** Yes, and that is the point. Managed AI Operations watches accuracy, safety and cost every day, re-tests the system on your own cases whenever anything changes, proves a new model before it reaches your users, and reports every month against the number agreed at the start. An engagement that ends at handover is the most common way enterprise AI quietly stops working. ## Applied AI for enterprise (https://applicat.ai/enterprise) Applicat AI works directly with mid-sized enterprises to design, build and run AI systems on frontier models. Most engagements begin with a fixed-fee sprint of four to six weeks that puts a working system on your own data and ends in a written go or no-go. If it is a go, the same engineers build it, roll it out and keep it running. Nothing is handed to a second team, and what they build belongs to you. ### The problem - **The demo that never ships.** A convincing demonstration on sample data wins the steering committee, and then meets the integration queue, the permissions model and the risk register. Nothing in the demo was wrong. It was simply never built to go anywhere, and that gap between the demonstration and the deployment is where most enterprise AI value is lost. - **The model you picked is already behind.** Capability, price and limits move every few months. A system wired to one vendor and one prompt gets expensive to maintain, and upgrading it turns into a rebuild nobody has the appetite to fund twice. - **When it underperforms, everyone points somewhere else.** Consultancies advise. Software vendors sell licences. Integrators build to the specification they were handed. All three can be blameless while the system sits there doing very little, because none of them ever agreed to be accountable for the result. ### How we engage - **Applied AI Sprint** (Frame and Prove). Four to six weeks at a fixed fee. We spend the first week or two ranking your candidate processes by what they are worth and how buildable they are. Then we build the top one on your own data, inside your own environment, and score it against cases your team chooses. You end with a number for what it costs to run and a written answer on whether to carry on. Includes: Your processes ranked by value and feasibility; A working system on your real data, in your environment; Scores against cases your own team picked; What it costs to run, and how fast it answers; A written go or no-go, with the evidence behind it. Fixed fee. Scope agreed before we start. - **Applied AI Programme** (Build and Deploy). The proof becomes a system your operation can depend on. It connects to what you already run, respects who is allowed to see what, keeps a record of every decision it makes, and clears your security review before release. Then we roll it out with the people who will use it and count whether they actually do. Includes: Wired into your existing systems and identity provider; Who may see what, and what the system may not do; A test set that every future change has to pass; A documented package for your security review; Rollout, training and a supported first period. Priced by milestone against the agreed outcome. - **Managed AI Operations** (Run). A live system is never finished. Volumes shift, the data behind it changes, and the models underneath it are replaced every few months. We watch accuracy, safety and cost daily, re-test on your own cases whenever anything moves, prove a new model long before it reaches your users, and write you a report each month against the number agreed at the start. End it whenever you like. The system is yours and it carries on running. Includes: A daily watch on accuracy, safety and spend; Re-tested on your own cases after every change; New models proven on your work before they go live; Named engineers on call, with agreed response times; A written monthly report to your sponsor. Monthly, with response times in the contract. - **Forward-deployed engineering** (Any stage). Named applied AI engineers working inside your teams for a defined period, in your tools and your stand-ups. They are there to move several use cases at once, and to leave your own people able to do this work without us. Includes: Named engineers, agreed before you sign; A weekly review against the outcomes; A written plan for handing the capability across; Our accelerators and testing tools, used on your work. Team-based, minimum three months. ### Security, data and governance - **No licence to protect, so no model to push.** Applicat AI is founder-owned. No AI lab, cloud provider, consultancy or investment fund sits on the cap table, and we earn nothing from anyone else's software sales. Which model goes into your system is settled by testing the candidates on your own cases, and we can tell you a use case is not worth building without it costing us a thing. - **Your process knowledge does not travel.** The code, the data and the tests are yours in the contract. We commit in writing that what we learn about how your business works stays with your business, and none of your material trains a shared model. - **It runs on your accounts, in your environment.** Systems are deployed into your own Microsoft Azure, AWS or Google Cloud environment, or into a dedicated one under your control. Model access goes through enterprise endpoints where your data is not used for training. End the relationship and nothing has to be moved. - **Your security team reviews a finished package.** Identity and access, data handling, logging, what the system may and may not do, and the due diligence on every vendor in the chain: all of it is written up during the Build stage, before release. Your reviewers get something complete to approve or reject. Nobody is asked to sign off an intention. - **Data that cannot leave the country does not leave.** Residency, retention and sovereignty requirements are captured in the first two weeks, and they drive the choice of model, region and architecture from there on. Where nothing may cross a border, the system is built on open-weight models running inside that jurisdiction. - **A record you can hand to an auditor.** Every system keeps its test results, its change log and the points where a human decides. Internal audit, a customer due-diligence questionnaire and the incoming AI regulations all ask for the same evidence, and it is written as the system is built, while it is still cheap to write. ### Enterprise questions **What does an applied AI engagement with Applicat AI cost?** The Applied AI Sprint is one fixed fee, agreed in writing before anything begins. Programmes are priced by milestone against the outcome they are scoped to. Managed AI Operations is a monthly fee with response times written into the contract. We give indicative ranges in the first conversation, so nobody spends six weeks on a proposal that was never going to be affordable. **Can Applicat AI work inside our existing cloud and security controls?** Yes. Systems are deployed into your own Azure, AWS or Google Cloud environment, on your accounts and behind your identity provider. Model access goes through enterprise endpoints where your data is not used for training. Identity, logging and data handling are documented during the Build stage, so your security review reads a finished package with the answers already written down. **Do we need a data platform before we start?** No. The first stage works out what data and which integrations each candidate use case would need, and favours the ones that can succeed with what you already have. A great deal of valuable applied AI runs on the documents, tickets, emails and PDFs that have been sitting in your systems for years. If a use case genuinely needs a data platform first, we will say so and rank it below the ones that do not. **How is Applicat AI different from a strategy consultancy or a systems integrator?** We are accountable for the system once it is live. The engineers who frame the work build it, deploy it and then run it, and the monthly report is written against the number agreed at the start. A consultancy sells advice and an integrator sells hours; our commercial model is tied to a system that works in production. We are also owned by nobody in the AI supply chain, so the model choice is made by testing, and we are free to recommend that you do not build at all. **Which industries does Applicat AI serve?** The Applicat Method works the same way in any sector. It is built for processes made of documents, requests, decisions and data, which covers most of the operational work in financial services, hospitality, retail, manufacturing, professional services, telecommunications, property and the public sector. The industries page sets out the use cases it covers in each. ## Business Operating System (https://applicat.ai/bos) A business operating system is the layer a company builds and runs its own AI on: one governed way into its systems, one answer to who may see what, one route to the models, one place where applications and agents are built and released, and one record of what they did. Applicat AI builds it inside your own environment, on your own accounts, for your people to build on. ### Nobody sets out to build an estate of nine AI vendors AI arrives one tool at a time: a chatbot from one vendor, a document tool from another, an agent bolted onto a single process by a third. Each purchase is defensible on its own. Together they become an estate nobody can govern, upgrade or switch off. - **A copy of your company in every vendor.** Each tool indexes your policies, contracts and tickets into its own store. The same document lives in five places, and five answers may disagree. - **Permissions that are not your permissions.** Every tool arrives with its own idea of who may see what, usually a service account with wider access than any employee holds, because that made the demo work. - **No shared answer to what happened.** Ask why a decision came out the way it did in March and you get as many log formats as you have vendors, several of which kept nothing worth reading. - **Nowhere for anything to be retired.** No register, no named owner, no review date. A tool nobody uses keeps its access and its invoice. None of this is an argument against buying software. It is an argument about where the seams go. The parts that ought to be shared, the connections, the permissions, the record of what happened, are the parts each product keeps inside itself. ### Five layers, and three things that run across all of them The phrase sometimes means a framework of meetings and metrics. This is not that. It is a set of parts assembled inside your own environment, most of them standard and several of them bought, so that what every application needs is held once, by you. In the stack, from the systems upward: 1. **Connections and tools.** One governed route into the systems the business runs on, so no application holds its own copy of your credentials. Contains: Finance, CRM, service desk, documents, mail and the warehouse, connected once; Credentials brokered once, never embedded in an application; Reads and writes exposed as named, scoped tools, revocable in one place. 2. **Knowledge and retrieval.** Search across your own material, inheriting the permissions of the system it came from, so nobody sees what their login would not show them. Contains: Documents, records and tickets indexed from the connected systems; One refresh path, so an answer is not quietly six weeks old; A citation back to the source record on every answer. 3. **Identity and permissions.** Agents and applications are principals in the identity provider you already run, each with a named owner. No more service accounts created on a Friday to make a demo work. Contains: Least privilege, scoped to a task and bounded in time; Acting for a user, or on its own behalf, whichever the task requires; One place to change what an agent may do. 4. **Model gateway.** Every model is reached through one gateway, so the choice of model is settled by a test result and changed by configuration. Contains: Frontier and open models behind a single interface; Routing by task, with fallbacks when a provider degrades; Cost and rate limits per application and per team; Residency respected where data cannot leave a jurisdiction. 5. **The build surface.** The templates, components and review path that turn something working on a laptop into something allowed to run. Contains: Templates and components for the patterns you repeat; Environments and versioning, so every change goes through a path someone approved; The same surface for your engineers and for ours. Across every layer: 6. **Evaluation harness.** One place your test cases live, so a model upgrade is a test run across everything you have built. Contains: A suite of your own cases behind every application and agent; Regression gates: nothing is released until it has passed those cases again; Baselines kept, so a new model is scored on your work and not a public benchmark. 7. **Observability and audit.** One trail across everything that runs: what was asked, what was retrieved, which tools were called, which model answered and who approved it. Contains: Run-level logging of inputs, retrieval, tool calls, model and cost; Human decisions recorded at the checkpoints where they are required; One export for internal audit, customer due diligence and AI regulation. 8. **The register.** What exists, who owns it, what it costs and whether anyone still uses it. Sprawl is prevented by having somewhere for things to be retired. Contains: Usage, cost and test status against every entry; A review date, and a defined way to switch something off. ### Your people, mostly. That is the point A layer only we can build on is a dependency with better manners. What your own people put on it is safe because the layer makes it so, not because someone remembered to be careful. - **Your engineers.** They inherit the connections, the identity model, the gateway and the test suites, so a new application starts at what is specific to your business. - **Your analysts and operators.** The people who know the process build the first version from a template, inside limits the layer enforces. Engineering comes in when a template will not carry it further. - **Our engineers, at the start and on the hard ones.** We build the first systems, the ones carrying real consequence, and the shared components everything else reuses. The balance shifts to your team as the layer fills out. We report the split: the share of what runs on the layer that your own people built. It is meant to rise, and if it does not, the layer is not doing its job and we would rather say so than invoice around it. ### The layer is what a programme leaves behind Nothing is bought in advance. Each part is built the first time a real system needs it, which is why the sequence is the same ladder as every other engagement. - **Applied AI Sprint** (Frame and Prove). The first use case is proved on your real data, inside your own environment, and the parts of the layer it needs are built properly the first time. What exists afterwards: One working system, measured on your own cases; Tests built from your cases, with baselines and a written go or no-go. - **Applied AI Programme** (Build and Deploy). The second and third systems reuse what the first one needed, and that reuse is what turns a system into a layer. Your engineers build alongside ours. What exists afterwards: Shared connections, identity, model access and a build surface; The register and the audit trail, on your own accounts. - **Managed AI Operations** (Run). The layer is operated. New models are tested on your cases before they are adopted, and entries that no longer earn their place are retired. What exists afterwards: Model upgrades proven before they reach users; A register that is reviewed, with cost and quality reported monthly. A company that will only ever run one AI system does not need a business operating system, and we will say so. The layer earns its cost where a second and a third system would otherwise each arrive with their own connections, permissions and version of the truth. ### End the relationship tomorrow and it keeps running Ownership is a test: could your team, or another firm, pick this up and carry it on? From the first sprint, the answer is meant to be yes. - **The code and the infrastructure.** Repositories, infrastructure as code and configuration, in your accounts and under your source control from the first week. - **Your environment, your data.** The layer runs in your own Microsoft Azure, AWS or Google Cloud tenancy, and model access goes through enterprise endpoints that do not train on your data. - **The tests and the baselines.** The cases, the scoring and the baselines are yours. They let you judge the next model, or the next supplier, without taking anyone at their word. - **No licence of ours in the middle.** The layer is assembled from standard and open components. There is no Applicat AI runtime to license, no fee per seat, and nothing that stops working when a contract with us ends. One honest limit. The models, and most of the systems the layer connects to, are other people's software, so independence here means you can swap any of them. ### The lock-in that matters is the layer, not the model Lock-in arguments are usually about whose model you use, which is the easiest part to change. What is hard to change is everything around it: the connections, the permissions, the index, the record of what happened. Whoever holds that layer holds the account. No lab, cloud, consultancy or fund owns us, so we build it for you to own, including against ourselves. - Models are swappable because they sit behind one gateway and one set of your own test cases. - Nothing in it requires Applicat AI to keep existing for it to keep running. ### Questions about the business operating system **Is this a product I can buy from Applicat AI?** No. There is no Applicat AI platform to license and no version number to quote. A business operating system is built for one organisation, inside its own cloud environment, from standard and open components, and owned by it. We sell the engineering that builds it, and the operation of it afterwards. **How is this different from buying an AI platform product?** Several of the layers can be bought, and where a product is the right answer for your identity provider or your data platform we will use one and say so. The question is which parts should be held once by you, and which are fine to buy inside a product. **Do we have to replace the AI tools we already have?** Usually not, and never all at once. The first step is an honest register of what exists, what each thing can reach and what it costs. Some tools earn their place and come under the layer's identity and audit model. Others are retired once an application on the layer does the same job. **How long before we own something real?** The first components exist at the end of the first sprint, because they are built to serve a working system. A layer that two or three systems genuinely share is the output of a programme, and that date is agreed once the first proof has been measured. **Who runs it after it is built?** Whoever you decide. Managed AI Operations can run it with your team, your team can run it with our support, or you can take it entirely. The layer sits on your accounts under your source control, and the handover pack is a deliverable of the Build stage. End the contract and it keeps running. ## Why independent (https://applicat.ai/why-independent) Applicat AI is an independent frontier applied AI company. The founders own it: no AI lab, no cloud provider, no consultancy, no investment fund, and no software licence of our own to protect. So the model that ends up in your system is the one that won on your own cases, what we learn about your operation stays inside your business, and when the honest answer is that you should not build the thing, we are free to say so. ### What changed in 2026 - 2026-03-16: Accenture. Completed its acquisition of Faculty, the London applied AI company, bringing more than 400 AI-native professionals and the Frontier decision intelligence product into the consultancy. (Source: Accenture newsroom, https://newsroom.accenture.com/news/2026/accenture-completes-acquisition-of-faculty) - 2026-05-04: Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs. Announced a joint venture of around $1.5 billion to build an enterprise AI services firm that embeds engineers and Claude models into mid-sized companies, with private equity portfolio companies as the natural first market. (Source: TechCrunch, https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/) - 2026-05-11: OpenAI. Launched the OpenAI Deployment Company with $4 billion of initial investment from 19 firms led by TPG, and agreed to acquire Tomoro, an applied AI consultancy with around 150 forward-deployed engineers. (Source: OpenAI, https://openai.com/index/openai-launches-the-deployment-company/) - 2026-07-02: Microsoft. Launched the Microsoft Frontier Company, an operating business with a $2.5 billion commitment and 6,000 industry and engineering experts, to deploy AI inside enterprise operations. (Source: TechCrunch, https://techcrunch.com/2026/07/02/microsoft-launches-its-own-ai-deployment-company-with-2-5-billion-commitment/) - 2026-07-15: Ode with Anthropic. The Anthropic joint venture launched publicly as a standalone services firm, led by chief executive Chris Taylor, targeting mid-sized organisations across financial services, healthcare, retail, manufacturing and software. (Source: Business Wire, https://www.businesswire.com/news/home/20260715205134/en/Anthropic-Blackstone-and-Hellman-Friedman-Introduce-Ode-with-Anthropic-an-Enterprise-AI-Services-Firm) ### What it means for a buyer - **Who chooses the model.** A deployment team owned by an AI lab deploys that lab's models. That may well be the right answer for your problem, but nobody has tested whether it is. The strongest model for reading documents, for voice, for code and for long reasoning work is rarely the same one, and the order changes every few months. - **Where your knowledge ends up.** A team working inside your operation learns how it really runs: the exceptions, the workarounds, the judgment nobody has written down. Boards have started asking where that understanding goes when the engagement ends, and whether a firm that also serves your competitors should be the one holding it. - **Who they were built to serve.** The new deployment ventures are aimed at the Fortune 500, the FTSE 100 and private equity portfolios, with minimum engagements reported in the millions. Mid-sized enterprises have the same processes, the same regulators and the same need for systems that hold up in production. Almost none of that new capacity points at them. ### Our commitments - **We test the models on your work first.** In the first weeks your own cases go through several candidates: frontier models from Anthropic, OpenAI and Google, and the open models where data residency, cost or speed favour them. You see the scores. The one that goes into your system is the one that won, and it can change when a better model arrives. - **What you tell us stays yours.** The code, the data, the prompts and the test cases are yours in the contract. We commit not to reuse what we learn about your operation on another client, and nothing of yours goes into training a shared model. - **Nobody holds a share of us but the founders.** No outside investor holds equity in Applicat AI, and that includes the AI labs, the clouds, the consultancies and the funds now buying their way into this market. If it ever changes, this page will say so before we say it anywhere else. - **Sized for organisations that decide in a room.** Scope, pace and price are set for mid-sized enterprises, and we go to them directly. That is why the first thing we ask you to commit to is a fixed-fee sprint of four to six weeks, at the end of which you can walk away with what we built. - **Independent enough to say no.** The Prove stage ends with a written go or no-go. We sell no models and no licences, so a no costs us the work and nothing more, and it saves you from the expensive maybe. - **Accredited with the labs, exclusive to none.** We pursue accreditations with the AI labs and clouds our clients use, and we have never traded one for exclusivity. Accreditation tells you we know a platform; independence tells you we will still choose the right one. ### Choosing a deployment firm | Question | Lab-owned deployment companies | Cloud provider deployment units | Global consultancies | Applicat AI | | --- | --- | --- | --- | --- | | Who picks the model | The owning lab's models | The owning cloud's catalogue | Broad, shaped by alliances | Whichever model wins on your own cases, frontier or open | | Who owns the firm | An AI lab plus investment funds | The cloud provider | Publicly listed or a partnership | The founders, and nobody else | | Who they serve | Fortune 500, FTSE 100, PE portfolios | Existing enterprise cloud accounts | Large enterprise and public sector | Mid-sized enterprises, engaged directly | | What you commit to first | Reported minimums in the millions | Tied to a platform commitment | A programme of work | A fixed-fee sprint of four to six weeks | | Where your knowledge goes | Set by the AI lab's terms | Set by the cloud's terms | Firm-wide knowledge reuse is the model | Nowhere: you own the assets, we do not reuse your process knowledge | | What happens after go-live | Varies by engagement | Tied to platform consumption | A managed services arm takes it on | We stay on the system, with a written report every month | General patterns as reported publicly in 2026. Any specific engagement may differ; ask each firm the same questions and get the answers in writing. ### Questions about independence **Why choose an independent applied AI firm over an AI lab's deployment company?** Three things change. The model is picked by testing it on your work, because nobody here has a licence riding on the answer. Your process knowledge and everything built with it stay yours under contract. And the engagement is sized for your organisation, where the new deployment ventures start at reported minimums in the millions. Plenty of organisations will end up using both, with the independent firm running the testing and running the system after go-live. **Is Applicat AI a partner of OpenAI, Anthropic or Google?** We build on their frontier models through enterprise programmes, and we pursue accreditations with the AI labs and clouds our clients use. We have never traded an accreditation for exclusivity. Which model goes into a client's system is settled by running that client's own cases through the candidates. **Does independent mean small?** Independence is a structure, not a size. Applicat AI has a globally distributed engineering team and was founded by the team behind BroadVision Technologies, which has delivered enterprise IT for more than 26 years. Every engagement is staffed with named engineers, and the Run stage carries service levels you can hold us to. **What happens to what you learn about our operation?** It stays with you. You own the code, the data, the prompts and the test cases, in the contract. We commit not to reuse your process knowledge on another client, and nothing of yours goes into training a shared model. End the relationship and your team keeps everything it needs to carry on. ## Frontier log (https://applicat.ai/frontier-log) The Frontier log is Applicat AI's running record of what changes for organisations running AI in production: model releases, pricing shifts, the moves the AI labs and cloud providers are making into deployment services, and the research that alters how we work. Each entry says what happened, and what we think it means for a system that is already live. We add to it as the frontier moves. ### 2026-09-02: Meta prices consent to train at a twelve-fold discount, and puts the switch in a config file What happened: Muse Spark 1.3 arrived from Meta with a second model string beside it, muse-spark-1.3-contributor, priced at $0.10 per million input tokens against $1.25 for the standard model and $0.20 against $4.25 on output, in exchange for permission to use prompts and completions to train future Meta models. The standard model does not train on customer data. Choosing between the two means changing a model identifier in configuration; no contract term governs it. What it means in production: The discount matters less than where the decision now sits. When no-training is a contractual protection, changing it takes procurement and legal. When it is a model name in a config file, any engineer with commit access can move a workload onto a training tier, and no data loss prevention tool, gateway or posture manager will see it, because nothing else about the request looks different. So if you allow this family of models at all, pin the permitted model strings at the gateway and alert on everything else. The model identifier is now a data boundary. Govern it like one. Source: Tech Times, https://www.techtimes.com/articles/326714/20260904/meta-muse-spark-contributor-tier-hides-training-consent-where-security-tools-cannot-find-it.htm ### 2026-09-02: Gemini 3.8 Flash ships with a published date for its own price rise What happened: Gemini 3.8 Flash launched with two prices. Google set it at $0.75 per million input tokens and $3.75 per million output through 31 December 2026, rising to $1.50 and $7.50 on 1 January 2027. Alongside it came Gemini 3.8 Flash Cyber, a vulnerability detection and patching model gated through the Fairwind programme to governments, critical infrastructure operators and software maintainers, with no published price. What it means in production: A doubling with a date attached is more useful than a quiet increase, and it belongs in the business case where finance can see it. A workload whose economics only work at the introductory rate does not have economics, it has a grace period. Model the unit cost at the list rate, keep the architecture able to move the workload when the rate changes, and put the date in the diary. The Cyber variant is the more interesting half. Its capability is rationed by eligibility, which is a procurement question most organisations are not set up to ask. Source: Google, https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/ ### 2026-09-01: Anthropic cuts cache reads by three quarters, and requires thirty-day retention to use the model What happened: Anthropic shipped the same underlying model twice, as Claude Fable 5.1 and as Claude Mythos 5.1 under different safeguards, with Mythos available only to organisations vetted through its cyber and life sciences verification programmes and initially only in the United States. Cache reads fell from $1.00 to $0.25 per million tokens while input and output held at $10 and $50. The migration guide lists a run of breaking changes, among them the removal of forced tool choice, the removal of the option to disable thinking, and a requirement for thirty-day data retention that returns a 400 error for organisations on zero-data-retention terms. What it means in production: Read the retention requirement first. Organisations that negotiated zero data retention, which is common in financial services and healthcare, cannot use this model without an explicit exception, and that is a procurement conversation with a longer lead time than the upgrade itself. The rest is the ordinary cost of living at the frontier: a real reduction on cached reads, paid for in integration work. That work is cheap if your prompts, tool definitions and model strings are version-controlled in one place, and expensive if they are scattered through application code, which is the actual argument for keeping the model layer behind a gateway you own. Source: Anthropic migration guide, https://platform.claude.com/docs/en/models/fable-5-1/migration-guide ### 2026-08-28: Z.ai holds an open-weight model back for two weeks over its own cyber capability What happened: GLM-5.3 was announced on 14 August. Z.ai then held the weights back for around two weeks, giving access first to vetted security partners before releasing them publicly at the end of the month. The company reported 84.5 per cent on CyberGym against 77.2 for GLM-5.2, and 54.4 per cent on ExploitBench against 24.4, and said the gains that help defenders find weaknesses earlier carry risk for attackers once weights are public. What it means in production: Open weights are no longer a clean answer to data residency. A staged release means the version you can self-host may trail the hosted one, and the gap will be widest exactly where capability is most sensitive, which is often the capability that made self-hosting attractive in the first place. So plan a regulated workload around the date the weights actually land. And be clear internally that self-hosting moves the safety question inside your perimeter; it does not remove it. Source: BetaNews, https://betanews.com/article/zai-glm-5-3-cybersecurity-delay/ ### 2026-08-02: The AI Act's transparency obligations become enforceable What happened: Enforcement of the AI Act began, run by the European Commission's AI Office and national authorities, with transparency obligations taking effect the same day: interactive systems must tell people when they are dealing with AI, deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks. The Commission published a list of more than 180 organisations that had signed the accompanying code of practice. What it means in production: For most production systems this is a small change that becomes an expensive one if nobody owns it until an auditor asks. Put disclosure in the interface and in the logs, where it will survive the next redesign. The machine-readable marking requirement is the one that reaches into architecture: the mark has to be applied at the point content is generated, which for most estates means touching several systems. Budget a quarter for that. Source: European Commission, https://digital-strategy.ec.europa.eu/en/news/commission-starts-enforcing-ai-act-rules-and-new-transparency-requirements-2-august ### 2026-07-30: Forward-deployed engineers become the industry's scarcest talent What happened: Demand for forward-deployed engineers is projected to surge by the end of 2026, according to TechCrunch: the share of companies planning to hire them jumped from under ten per cent early in the year to around seventy per cent by the second quarter, against a pool of roughly 17,000 in the United States of whom only around 2,000 have elite-level expertise. What it means in production: Capacity is now the constraint on getting AI into production, and it is a hiring constraint. Organisations that cannot hire this profile will buy it as a service, they will judge the firms they buy it from on the depth of the bench, and the methods that survive will be the ones that hand capability to the client's own people as they go. Source: TechCrunch, https://techcrunch.com/2026/07/30/forward-deployed-engineers-are-the-ai-industrys-latest-talent-obsession/ ### 2026-07-15: Ode with Anthropic launches as a standalone services firm What happened: Ode with Anthropic launched as an enterprise AI services firm under chief executive Chris Taylor, introduced by Anthropic, Blackstone and Hellman & Friedman and backed by a consortium including Goldman Sachs, General Atlantic, Apollo, GIC and Sequoia. It is aimed at mid-sized organisations across financial services, healthcare, retail, manufacturing and software. What it means in production: A second frontier AI lab now sells deployment as well as models. That collapses two procurement questions into one: you cannot ask who deploys the system without also asking whose models they have an interest in selling. We expect model-neutral evaluation to become a standard procurement requirement. Source: Business Wire, https://www.businesswire.com/news/home/20260715205134/en/Anthropic-Blackstone-and-Hellman-Friedman-Introduce-Ode-with-Anthropic-an-Enterprise-AI-Services-Firm ### 2026-07-02: Microsoft commits $2.5 billion and 6,000 people to the Frontier Company What happened: With 6,000 industry and engineering experts, the Microsoft Frontier Company opened as an operating business that deploys AI inside enterprise operations, positioned as the largest outcome-driven engineering organisation in the industry. Early customers include London Stock Exchange Group and Unilever. What it means in production: For organisations already committed to Azure and Microsoft 365, deployment capacity has now arrived inside the vendor relationship, carrying the platform dependency that implies. For everyone else, and for mid-sized enterprises in particular, the gap is unchanged, and model-agnostic architecture is the hedge either way. Source: TechCrunch, https://techcrunch.com/2026/07/02/microsoft-launches-its-own-ai-deployment-company-with-2-5-billion-commitment/ ### 2026-05-19: Gartner puts 2026 AI spending at $2.59 trillion, with $585 billion on services What happened: Gartner's forecast puts worldwide AI spending at $2.59 trillion in 2026, up 47 per cent, with AI services at around $585 billion. Gartner described 2026 as the inflection year in which enterprises expand embedded generative AI and new agents across multiple workflows, while noting limited appetite for disruptive change in favour of tactical initiatives. What it means in production: The money is moving from experimentation into services and operations. What buyers say they want is tactical, well bounded and attached to a number, which favours anyone whose method starts from a narrow value case and an evaluation set. The premium sits with firms that can run many small systems well. Source: Gartner, https://www.gartner.com/en/newsroom/press-releases/2026-05-19-gartner-forecasts-worldwide-ai-spending-to-grow-47-percent-in-2026 ### 2026-05-11: OpenAI launches the Deployment Company and acquires Tomoro What happened: The OpenAI Deployment Company was set up as a majority-owned business, with $4 billion of initial investment from 19 firms led by TPG and including Bain & Company, Capgemini and McKinsey. On the same day OpenAI agreed to acquire Tomoro, a London and Edinburgh applied AI consultancy with around 150 forward-deployed engineers and clients including Tesco, Virgin Atlantic and Supercell. What it means in production: The model company is now a services company. Traditional consultancies and integrators saw their shares fall on the news. Capacity to deploy OpenAI models went up that day, and so did the value of having someone independent run the head-to-head. Source: OpenAI, https://openai.com/index/openai-launches-the-deployment-company/ ### 2026-03-16: Accenture completes its acquisition of Faculty What happened: Faculty, the London applied AI company founded in 2014, became part of Accenture on completion of the acquisition, taking more than 400 data scientists and AI engineers and the Frontier decision intelligence product with it. Faculty's founder Marc Warner became Accenture's chief technology officer. What it means in production: The independent applied AI bench in Europe got smaller on the day the market needed it most. Consolidation into a global consultancy brings scale and platform alliances, and it moves the firm's incentives along with them. Outside the largest accounts, independent capacity is now scarce. Source: Accenture newsroom, https://newsroom.accenture.com/news/2026/accenture-completes-acquisition-of-faculty ### 2025-08-19: MIT NANDA: 95 per cent of enterprise generative AI pilots show no P&L impact What happened: The GenAI Divide: State of AI in Business 2025, published by MIT's NANDA initiative, found that around 95 per cent of enterprise generative AI pilots produced no measurable impact on profit and loss. It also found that deployments built with external specialists succeeded about twice as often as internal builds, and that the tools that stalled could not retain feedback or adapt to context. What it means in production: This is the report that named the problem applied AI has to solve. Its findings are the basis of the Applicat Method: narrow processes, proofs built in production conditions, evaluation over opinion, and an operating model for the years after go-live. Source: MIT NANDA report (PDF), https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf ### Questions about the frontier log **How often is the Frontier log updated?** Whenever something moves that changes how AI systems should be built or run in production: a model release, a pricing change, a major market move or a piece of research. There is no publishing schedule, because the frontier does not keep one. **Why does a services company publish market commentary?** Because our clients have to make decisions in this market, and our method depends on reading the frontier accurately. Publishing the log keeps our own reading honest, and gives the people we work with a shared reference for the conversations that follow. ## Applied AI readiness check (https://applicat.ai/readiness) The Applied AI Readiness Check is ten questions about the five things that decide whether an AI system ever reaches production: a value case worth the effort, the data, delivery, governance, and who owns it afterwards. It takes three minutes, runs entirely in your browser, and ends with a profile and a recommended next step. Nothing is sent anywhere unless you choose to send it. Questions: 1. (Value) Can you name a specific process where AI would change a number leadership already tracks? 2. (Value) Is there a named sponsor who owns that number and can make time for the work? 3. (Data) Does the data for that process already exist in documents, tickets, emails or systems you run? 4. (Data) Are the systems involved reachable through supported APIs or connectors? 5. (Delivery) Could a four to six week proof run on real data inside your security perimeter? 6. (Delivery) Do you have people who can work alongside an engineering team weekly for the duration? 7. (Governance) Do you know which actions in the process a person must approve, and which can be automated? 8. (Governance) Are data residency, retention and model-provider rules defined for AI use? 9. (Operations) Who would watch quality, cost and incidents after go-live? 10. (Operations) How would you decide whether a new model release should replace the current one? Profiles: - **Exploring** (0 to 6 of 20). The interest is real and the foundations are not there yet: no measured value case, uncertain access to the data, no governance baseline. Starting a build now would produce a demonstration and nothing you could put in front of a customer. Next step: Run the Frame stage on its own. In one to two weeks it gives you a ranked list of what is worth building, a straight answer on whether the data and the systems can be reached, and the governance requirements: exactly the list this check found missing. - **Pilot-bound** (7 to 12 of 20). You could start a pilot tomorrow, and that is the risk. Most organisations at this level end up with a demonstration that cannot pass security, integration or ownership questions later. Next step: Choose one use case with a number attached to it, settle the access and approval questions before anyone builds anything, and run an Applied AI Sprint so the first working version sits inside your own environment from day one. - **Production-ready** (13 to 17 of 20). Value cases, data access and sponsorship are in place. What is usually still missing is governance detail and a decision about who owns the system once it is live. Next step: Go straight to an Applied AI Sprint on the highest-value case. Agree what happens after go-live at the same time: who watches quality and cost, what gets re-tested on every change, and how a new model gets approved before it reaches users. - **Scaling** (18 to 20 of 20). You have the foundations to run a portfolio, not a project. The challenge now is throughput: how many systems can reach production each quarter without quality or cost drifting. Next step: An Applied AI Programme with forward-deployed engineers across a portfolio of use cases, and Managed AI Operations to keep every system tested, governed and reported on as the frontier moves underneath it. ### Questions about the readiness check **Is my answer data stored or sent anywhere?** No. The readiness check runs entirely in your browser. Nothing leaves your device unless you choose to email your result to us using the button at the end. **What is a good readiness score?** Thirteen or more out of twenty means an Applied AI Sprint can realistically put a production-grade system on your real data in four to six weeks. Below that, the fastest route is usually a short Frame stage that closes the specific gaps this check found. The total matters less than which of the five dimensions came out lowest. ## Solutions (https://applicat.ai/solutions) Six kinds of system, each one pointed at work that is costing you time or money today. Every one is built on your own data, runs in your own environment on your own accounts, and belongs to you in the contract. And once it is live somebody is still watching it: accuracy and cost checked, every change re-tested on your own cases, and a written report each month against the number we agreed before we started. ### Agentic workflow automation (https://applicat.ai/solutions/agentic-workflow-automation) Agentic workflow automation hands a whole process to software agents: they read what arrives, look up what they need across your systems, take the next action, and stop at the points where a person should decide. The processes worth doing this to are the ones that cross four systems and three inboxes before anyone can close them, where the exceptions eat more of the week than the straightforward cases ever do. We map the sequence and the stopping points with the people who run it today, build it against your systems of record, and stay on it afterwards: every action logged, every change re-tested on your own cases, and a monthly report written against the number agreed at the start. Outcomes: Cases closed in minutes that used to sit overnight; Volume goes up without the headcount going up with it; Every action logged, attributable and reversible; Your people spend the day on the exceptions. Use cases: - **Finance operations.** Invoice matching, accruals, intercompany reconciliation, expense review and the month-end tasks that always run late. The controller still signs. - **Procurement and suppliers.** Requisitions triaged, supplier checks run, renewals prepared, and purchase order exceptions chased before they turn into a phone call from the supplier. - **Onboarding.** A new employee or a new customer needs documents collected, identities verified, accounts requested and somebody kept informed. That work sits across HR, IT and the business, which is exactly why it falls between them. - **Claims and cases.** Intake, completeness checks, policy lookup, a drafted recommendation and routing. The adjudication stays with the person who is paid to make it. - **Compliance and controls.** Evidence gathered, control testing prepared, attestations chased, and the first draft of a regulatory return assembled. How it works: 1. Map the process the way it actually runs, workarounds and exceptions included. That map is usually the first thing a client keeps. 2. Agree what each agent may do, what it may touch, and where it has to stop and ask a person. 3. Build the integrations into your systems of record, with the narrowest access that does the job. 4. Score it on cases you have already closed. You take the go-live decision against that score, and every change afterwards has to clear it again. Good fit: Best where the rules are clear, several systems are in the loop, and the exceptions are absorbing skilled people you would rather have doing something else. #### Questions about agentic workflow automation **What is the difference between agentic automation and RPA?** Robotic process automation follows a fixed script and breaks the moment an input looks different. An agentic system uses a frontier model to read what has arrived, work out the next step and cope with variation, inside a fixed set of tools, permissions and checkpoints agreed in advance. The practical difference shows up in the exceptions. RPA hands all of them back to a person. An agentic system handles most of them and escalates the rest with its reasoning attached. **How do you stop an agent taking the wrong action?** Four things hold it. It can only use the tools you have given it and only reach the systems you have opened to it. Consequential actions stop at a human checkpoint. Outputs are checked before they are acted on. And every change, including a new version of the underlying model, has to pass the tests built from your own cases before it reaches production. Every action is logged, so when something does go wrong you can see exactly what happened and undo it. **Can agents work across our ERP, CRM and ticketing systems?** Yes. We integrate through supported APIs and connectors for platforms such as SAP, Microsoft Dynamics, Sage, Salesforce, ServiceNow and HubSpot, and through secure automation where no API exists. Each agent is given its own identity and the narrowest set of permissions that lets it do the job, so what it can reach stays a decision your security team makes, and can revoke. ### Document intelligence (https://applicat.ai/solutions/document-intelligence) Document intelligence puts frontier models on the contracts, invoices, claims, forms and scans that arrive every day: pulling out the fields, sorting and routing, checking one document against another, and holding back anything it is unsure about for a person to look at. Every value it produces points back to the page it came from. Before any of it touches a live process we measure how often it is right, field by field, on your own documents with the bad scans left in, and you decide against that number whether it goes live. Outcomes: Hours of reading taken out of every case; Accuracy measured field by field on your own documents, before go-live and after it; Every value traceable back to the page it came from; Reviewers see the documents that genuinely need a reviewer. Use cases: - **Contracts and obligations.** Clauses extracted, deviations from your standard terms flagged, and renewal dates and obligations tracked across an estate nobody has read end to end. - **Invoices and remittances.** Line-level extraction, three-way matching, and exceptions routed to the person who can clear them. - **Claims, applications and forms.** Mixed-quality scans, photographs and the occasional handwritten page, classified, checked for completeness and turned into fields somebody can work with. - **Regulatory and technical documents.** Requirements extracted, two versions of a standard compared to show exactly what changed, and evidence mapped for an audit. - **Onboarding and due diligence packs.** Identity documents, certificates, registers and proofs of address checked against each other and against what the file already says. - **Correspondence.** What the message is about, what it refers to, who should have it, and a draft reply for them to work from. How it works: 1. Take a real sample, including the scans nobody wants to open, and set the baseline on that. 2. Design the fields and the review rules with the people who do the work today. 3. Build the pipeline: confidence thresholds, a review queue for whatever falls below them, and a link back to the page behind every value. 4. Measure precision and recall field by field. You take the go-live decision against those figures, and they keep being measured once it is live. Good fit: If thousands of documents a month pass through people who are mostly transcribing them, this is the shortest route to a number that moves. #### Questions about document intelligence **How accurate is AI document extraction?** It depends on the document, the field and the model, which is why we will not quote a figure before we have seen your documents. During the Prove stage we measure precision and recall field by field on your own material, bad scans and edge cases included, and you take the go-live decision against the accuracy your process actually requires. Some fields have to be right every time. Others only have to beat what happens at five on a Friday afternoon. **Do you need our documents to train a model?** Usually not. Frontier models do this well with a clear schema, good instructions and a handful of examples. We fine-tune only where the measurements show it earns its place, and your documents are never used to train a shared model. **Can it handle handwriting and poor scans?** Often, yes. Frontier multimodal models cope with a wide range of handwriting and scan quality, and they report when they are unsure. Where confidence is low the document goes to a person, and that case is added to the test set so every later version is measured on it. ### Customer and service operations (https://applicat.ai/solutions/customer-and-service-operations) Requests arrive all day by email, chat, phone and portal, and most of them are versions of the same few dozen questions. This is the system that sorts what comes in, answers the routine cases from your order systems and your policy documents inside limits you set, drafts the reply and the background for whoever handles the rest, and hands over cleanly the moment a customer asks for a person. Quality review then runs across all of it, every call and every thread, which is a good deal further than hand sampling ever reaches. Outcomes: Routine requests answered without anyone joining a queue; First response and resolution times down across every channel; The same question gets the same answer, from current policy; Quality reviewed on all of it, every call and every thread. Use cases: - **Triage and routing.** What it is about, how urgent it is and who should have it, decided the moment it arrives, so nothing waits for the morning queue review. - **Answering the routine cases.** Order status, account changes, bookings, password and access requests, completed end to end inside limits you set. - **Help for the person on the call.** A suggested answer with its source, a drafted reply, the account history summarised, all inside the tools your team already has open. - **Quality and coaching.** Every interaction scored against your own framework, surfacing the compliance flags and the coaching points a supervisor working from a sample would never see. - **What customers keep telling you.** Themes and root causes pulled out of the whole body of interactions while there is still time to act on them. How it works: 1. Read the history first. The volumes and the request types decide what is worth automating and in what order. 2. Ground the system in your policies, your product data and your systems of record, with the source shown alongside the answer. 3. Set the limits with your operations leads: what it may resolve on its own, what it must hand over, and how that handover reads to the customer. 4. Containment, accuracy and what the customer actually experienced all get measured before go-live, and again every month after it. Good fit: Suited to operations taking tens of thousands of interactions a month across several channels, where the same questions keep coming back. #### Questions about customer and service operations **Will customers know they are talking to AI?** Yes. Where regulation or your own policy requires disclosure it is disclosed, and we recommend disclosing everywhere else as well. The system is built to hand over to a person with the full history whenever a customer asks or whenever it is not confident, and that handover is designed before anything goes near a customer. **Does this replace our contact centre platform?** No. We work with what you already run, integrating with platforms such as Genesys, Zendesk, Salesforce Service Cloud, ServiceNow, Freshworks and Microsoft Dynamics. The intelligence goes into the tools your team already has open, which is also the quickest way to get it used. **What containment rate should we expect?** That depends entirely on the mix of requests you receive, so anyone quoting a figure before looking at your data is quoting an industry average. During the Prove stage we measure containment on your own historical interactions and the business case is built on that number. It is also the number the monthly report is written against once the system is live. ### Knowledge and decision support (https://applicat.ai/solutions/knowledge-and-decision-support) What an organisation knows tends to sit in three places at once: a policy library nobody has opened since the last revision, a shared drive that goes back further than anyone still working there, and the heads of four people who are busy. Knowledge and decision support systems answer questions from that material in seconds and attach the passage the answer came from. Access follows the permissions you already have, so nobody receives an answer assembled from a document they are not allowed to open. For heavier work the same grounding produces a structured brief, with the policy, the precedent and the live numbers in one place for whoever has to sign. The pattern underneath is retrieval-augmented generation, usually shortened to RAG, and the tests that keep it honest are built from real questions your own experts have already answered. Outcomes: An answer in seconds, with the passage it came from attached; One policy applied the same way in every team and every country; It tells you when the sources do not cover the question; Expert time kept for the questions that need an expert; Nobody sees an answer built from a document they cannot open. Use cases: - **Policy and procedure.** Grounded answers on HR, finance, compliance and operational policy, versioned so the answer changes on the day the policy does. - **Engineering and field knowledge.** Manuals, standards, drawings and twenty years of incident reports, answerable by the technician standing in front of the machine. - **Bids and proposals.** Drafting from approved content and past bids, with a compliance check against what the tender actually asked for. - **Research and analysis.** Synthesis across reports, filings and internal work, with structured output and every source tracked. - **Decision briefs.** Policy, precedent and live numbers pulled into one structured recommendation for a credit, pricing, risk or approval decision. How it works: 1. List the sources, who owns each one, how often it changes and who is allowed to see it. This step surfaces more than it sounds like it will. 2. Build retrieval that respects those permissions and returns the paragraph, with the document standing behind it. 3. Collect real questions from the people who ask them, and the answers your own experts consider correct. 4. Before go-live, and again whenever the content or the model changes, we check how often the answer actually came from the sources. Good fit: Right when the expertise sits with a handful of people, or is spread across so many systems that nobody can answer anything without asking somebody. #### Questions about knowledge and decision support **What is retrieval-augmented generation?** Retrieval-augmented generation, usually shortened to RAG, means the system looks up the relevant passages in your own content at the moment a question is asked and hands them to the model as context. The answer is then built from your sources and can be cited back to them, which is what makes it checkable by the person reading it. **How do you stop the assistant making things up?** By giving it the passages it has to answer from, requiring a citation, and designing it to say that the answer is not in the sources instead of producing something plausible. Then we measure it: a set of real questions with answers your own experts have approved, run before go-live and again whenever the content or the model changes. Where a question falls outside the sources the system says so and points the person at someone who knows. **Does it respect our document permissions?** Yes. Retrieval is permission-aware, so a person only ever receives an answer built from content they are already entitled to open, and every access is logged. If someone cannot see the document today, they cannot see it through the assistant either. ### Data and analytics agents (https://applicat.ai/solutions/data-and-analytics-agents) Data and analytics agents let people ask questions of governed data in plain words, then write the query, check it, and show the definitions behind the number they return. They are built on your semantic layer, so what comes back agrees with the figures your finance and data teams already publish, and we test that agreement on questions your data team has answered before anybody else is given access. Outcomes: Questions answered the same afternoon they are asked; One set of definitions behind every number; Recurring reports drafted, with the variances already explained; A metric moving the wrong way flagged before the monthly review. Use cases: - **Questions in plain words.** Asked of the warehouse or lakehouse, answered with the query shown and the definitions it used set out beside the number. - **Reports and commentary.** Board packs, operational reviews and client reports drafted, including the paragraph that explains why the numbers moved. - **Monitoring and alerting.** Metrics watched continuously, with an account of what changed and where, sent to whoever can act on it. - **Data quality.** Checks written, run and triaged, with proposed fixes going to the owner of the data. - **Planning support.** Assumptions gathered, models re-run and results explained, through a planning cycle that used to be a fortnight of spreadsheets. How it works: 1. Start from your metric definitions. Where they do not exist yet, agreeing them is the first piece of work, and it is worth doing anyway. 2. Build the agent to write a query, validate it, and show its working. 3. Test it on questions your data team has already answered, compare, and then put it where people already are: Microsoft Teams, Slack, Power BI, the browser. Good fit: Worth doing where there is a warehouse or a lakehouse already, and a queue of business questions that never reaches the top of the backlog. #### Questions about data and analytics agents **Will the agent give different numbers from our dashboards?** Not if it is built on the same definitions. We build it against your semantic layer so it uses the metrics your reporting already uses, and we test agreement with published numbers before anyone gets access, then again whenever a definition changes. Where the agent and a dashboard do disagree, that is usually a definition problem you had already, now visible. **Which data platforms do you support?** Snowflake, Databricks, Google BigQuery, Microsoft Fabric and Azure Synapse, Amazon Redshift, SQL Server and PostgreSQL, alongside Power BI, Tableau and Looker for presentation. **Is our data sent to the model provider?** Only the minimum context needed to answer the question, through enterprise endpoints that do not train on your data, and never anywhere your residency rules forbid. Where nothing may leave at all, open models running inside your own environment on your own accounts are a workable option, and we will tell you plainly what that choice costs you in capability. ### Custom AI applications (https://applicat.ai/solutions/custom-ai-applications) Sometimes nothing on the market does the thing, because the capability you need sits in the model itself. Custom AI applications are the products and internal tools built on that: a customer-facing feature, an assistant for a piece of regulated work, or a platform your own teams build on top of. We design the architecture so the model underneath can be changed when a better one arrives, prove the idea on your cases before it is built out, and hand over an operating model that says who watches quality, who approves an upgrade and who owns the bill. Outcomes: A product on the frontier without hiring a research team; An architecture that survives the next generation of models; Someone accountable for quality, latency and cost once it is live. Use cases: - **AI-native products.** Customer-facing features where the model is the product, from specialised assistants to generation and analysis tools. - **Internal platforms.** Shared retrieval, agents, testing and governance, so your teams build on one foundation and not six. - **Assistants for expert work.** Underwriting, clinical administration, legal review, engineering design. Work where being roughly right is worse than being no help at all. - **Multimodal systems.** Images, audio, video and documents understood together, for inspection, monitoring and media work. - **Choosing the model.** An independent read on which frontier or open model wins on your cases, what it costs to run, and how you would move if that changes. We have no licence riding on the answer. How it works: 1. Agree the outcome and the test that will prove it, before anyone writes code. 2. Choose the model by running the candidates on your cases, and build the model layer so it can be swapped later. 3. Build the application itself: orchestration, guardrails, monitoring and cost controls. 4. Prove it on real users, in a narrow slice, with a way back. 5. Hand over the operating model: who watches quality, who approves a model upgrade, who owns the bill. Good fit: The right call when you have a clear product idea and a real user at the end of it, and nobody at hand who has shipped on frontier models before. #### Questions about custom ai applications **Do you fine-tune models?** Where the measurements show it earns its place. Most applications do better on a frontier model with well-designed context, retrieval and tools than on a fine-tuned smaller one. Fine-tuning becomes worth it for narrow repetitive tasks, for hard latency or cost targets, or when an open model has to run inside your own environment because nothing may leave it. **Who owns the intellectual property?** You own the application, its code, its prompts and the tests built from your own cases. Applicat AI keeps its own accelerators and tooling, and licenses those to you for the life of the system. End the relationship tomorrow and what we built keeps running, with your team holding everything they need to change it. **What happens when a better model comes out six months in?** The model layer is designed so a new one can be dropped in. We run it against the tests built from your cases, compare quality, latency and cost with whatever is running now, and show you the comparison. It only reaches your users if it wins. That work is part of Managed AI Operations, and it is the reason a system built this year is still the right system in three years. **Can you take over an application another vendor built?** Yes, and we start by measuring what it actually does today, because that is usually unknown. Then we stabilise it, add the testing and the monitoring that were missing, and move it onto an architecture where the model underneath can be changed. Sometimes the honest answer after that assessment is that rebuilding costs less than rescuing, and we will put that in writing. ## The Applicat Method (https://applicat.ai/method) The Applicat Method is Applicat AI's five-stage delivery model for applied AI: Frame, Prove, Build, Deploy and Run. Each stage produces something you can look at, and each ends in a decision you make on the evidence. Any one of the five can be the last, which is the point of drawing it this way. Most enterprise AI does not fail loudly. It sits in a pilot nobody is willing to kill. ### Five stages 1. **Frame** (1 to 2 weeks). We find the places where AI could move a number your leadership already watches, then rank them by what they are worth, how buildable they are and whether the data exists. You come out with a shortlist, a value case for each candidate, and an agreed description of what success would look like. Outputs: Your processes ranked by value and feasibility; What data and integrations each one would need; The number each candidate is meant to move; Risk, residency and governance requirements. 2. **Prove** (3 to 4 weeks). We build the top candidate for real: your own data, your own environment, the permissions that actually apply. It is then scored against cases your team picks, and measured for what it costs and how fast it answers at the volumes you expect. The stage ends in a recommendation, and no is a normal outcome. Outputs: A working system on real data and real integrations; A test set drawn from your own cases, with baseline scores; Cost and speed at the volumes you expect; A written go or no-go, with the evidence. 3. **Build** (4 to 10 weeks). The proof becomes something the business can depend on. It connects to the systems you already run, respects who is allowed to see what, keeps a record of every decision it makes, and refuses what it should refuse. Security review happens in this stage, before release. From here on, no change reaches a user until it has passed the tests, which is what an evaluation suite is for. Outputs: Production architecture and integrations; Permissions, audit trail and limits on what it may do; Every change re-tested automatically before release; Security and compliance documentation. 4. **Deploy** (2 to 4 weeks). We roll out with the people who will actually use it, agree the points where a person has to decide, train the teams and count usage from the first week. A system nobody opens is worth nothing, so adoption is measured and reported like any other number in the business. Outputs: Rollout plan and human checkpoints; Training and playbooks for the teams; Adoption and value reporting; A supported first period after go-live. 5. **Run** (Ongoing). Volumes shift, the data behind the system changes, and the models underneath it are replaced every few months, so a live system left alone quietly gets worse. We watch accuracy, safety and cost daily, re-test on your own cases whenever anything moves, prove a new model before it reaches your users, and write a report each month against the number agreed at the start. Outputs: Accuracy, safety and cost watched daily; Model upgrades proven before they reach users; Continuous testing and improvement; A monthly report against the agreed number. ### Principles - **The proof runs where the real thing will run.** Proofs are built on your real data, in your own environment, with the permissions that genuinely apply. There is then no gap between what was demonstrated and what can be deployed, and that gap is where most pilots quietly die. - **Nothing ships until it passes its tests.** Every system carries a set of tests written from your own cases. They decide the go-live, they decide whether a model upgrade happens, and they catch a regression before your users meet it. That set is the evaluation suite, and it stays with you. - **People keep the decisions that carry consequences.** Agents take the volume. We design the points where a person decides, and we keep those points fast, because a checkpoint that costs somebody a day is a checkpoint they learn to route around. - **Your data never leaves your control.** Systems run on your own cloud accounts, or in a dedicated environment you control. Your material is not used to train shared models, and residency is settled at the Frame stage, before any architecture is fixed. - **When a better model arrives, you can take it.** Model capability moves every few months. We keep the model behind an interface with its own tests, so swapping one for another becomes a test run and a decision made on the scores. - **The people who scope it are the people who run it.** No hand-off from a strategy team to a delivery team, and no gap between the advice and the consequence. The engineers in the first workshop are the ones you call at go-live. ### Questions about the method **How long does it take to get an AI system into production?** A working system on your own data takes four to six weeks: one to two weeks of Frame, then three to four of Prove. How long it then takes to go live depends on your integrations and your governance, so Build and Deploy are scoped once the proof has been measured. Frame, Prove and Build together come to eight to sixteen weeks, which is the window a first production release falls into, and we put a date on it once the proof has been measured. **What is the difference between a proof and a pilot?** A pilot is usually a demonstration on sample data, outside the production environment, built to be watched rather than used. A proof is a working system on your real data, inside your own environment, scored against your own cases and built to the standard of something that could go live. The difference shows at deployment: a proof already has the integrations, the permissions and the test results, so the path to production is short. **Do you work with our existing IT and data teams?** Yes, from the first week. Our engineers work inside your tools and your stand-ups, alongside the people who will own the system afterwards. Handing the knowledge across is built into the Build and Deploy stages week by week, and you decide how much of the Run stage stays in-house. **What happens if the proof shows the use case does not work?** We say so in writing, with the measurements behind it, and we recommend against proceeding. The Frame stage produces a ranked shortlist, so there is normally a next candidate ready to test. A clear no in week five is a far cheaper outcome than a slow maybe that runs for a year, and we have no licence sale riding on the answer, which is part of why we are able to give it. ## Industries (https://applicat.ai/industries) The five stages, the testing discipline and the way we run a system after go-live are the same whether the process belongs to a bank, a hotel group or a manufacturer. What changes is the domain knowledge your people already hold, the systems we connect to and the regulation we design for. Whatever the sector, the system runs on your own accounts and what we build stays yours. Below are the use cases the method covers, sector by sector. ### Financial services Banks, insurers, asset managers and payment businesses operating under close regulation with high document volumes and strict audit requirements. - Know-your-customer and onboarding document review - Claims intake and adjudication support - Credit memo and decision brief drafting - Complaints triage and root-cause analysis - Regulatory reporting preparation and control testing ### Hospitality and leisure Hotel groups, resorts, estates and leisure operators with multi-property operations and guest experience at the centre of the business. - Guest request triage and resolution across channels - Reservation, group and event correspondence handling - Revenue and operations reporting with commentary - Supplier invoice matching across properties - Brand standard audits from inspection reports and photographs ### Retail and consumer Retailers, franchisors and consumer brands running high-volume customer operations and complex supply chains. - Customer service resolution and agent assist - Product content generation and catalogue quality - Returns, refunds and dispute handling - Store operations reporting and anomaly alerts - Supplier and franchisee correspondence ### Manufacturing and industrial Manufacturers, mining and energy operators and industrial service companies with engineering knowledge locked in documents and experienced people. - Maintenance and engineering knowledge assistants - Quality and incident report analysis - Procurement and supplier document processing - Health, safety and environmental reporting - Tender and specification review ### Professional services Legal, accounting, advisory and engineering firms where expert time is the product and knowledge reuse is the margin. - Contract and document review at scale - Proposal and bid drafting from approved content - Engagement knowledge capture and reuse - Time, billing and work-in-progress analysis - Research and regulatory monitoring ### Telecommunications and media Operators, broadcasters and media groups with very high interaction volumes, complex products and content operations. - Customer care resolution and retention support - Order fall-out and provisioning exception handling - Content metadata, compliance and rights processing - Network and service ticket triage - Subscriber analytics and churn commentary ### Property and real estate Developers, estate managers and property services businesses handling leases, tenants, suppliers and compliance across portfolios. - Lease abstraction and obligation tracking - Tenant and resident request handling - Supplier invoice and works order processing - Compliance evidence collection across sites - Portfolio reporting with commentary ### Public sector and not-for-profit Government bodies, agencies and charities where citizen and beneficiary service, transparency and cost discipline all matter. - Citizen and beneficiary enquiry handling - Application and grant assessment support - Case file summarisation and routing - Policy and procedure assistants for staff - Freedom of information and records processing ### Questions about industries **Does Applicat AI specialise in a particular industry?** No, and that is deliberate. The domain knowledge comes from your teams and from our engineers' experience across sectors. What makes a system specific to your business is that it is tested on your own cases, with your own documents, your own exceptions and your own definition of a right answer. **Can you work in regulated industries?** Yes. Financial services, healthcare administration and public sector engagements are designed with the relevant regulation from the Frame stage: data residency, auditability, human oversight and model governance are built into the system rather than added afterwards. **My industry is not listed. Can you still help?** Almost certainly. The solution areas, from document intelligence to agentic workflow automation, apply wherever work is made of documents, requests, decisions and data, which is most work. Tell us the process you have in mind and we will tell you whether it is worth doing. ## Insights (https://applicat.ai/insights) Evidence-led articles from the Applicat AI team on getting AI into production: definitions, methods, model selection, governance and the economics of running a system after go-live. Written for the leaders and engineers who have to make it work. ### The model companies are now the services companies. What that means for you (https://applicat.ai/insights/the-model-companies-are-now-the-services-companies) By Luka Kokot, Founder and Chief Executive, Applicat AI. Published 2026-09-18, updated 2026-09-18. Within five months of 2026, the two leading AI labs and the two largest clouds each launched a deployment business, and the largest consultancy bought Europe's best-known applied AI firm. The model layer is commoditising and the margin is migrating to deployment. For buyers, "which model" and "who deploys it" have stopped being separate questions unless the firm doing the deploying is independent. Here is what happened, why, and how to choose a deployment firm now. Key takeaways: - OpenAI (May), Anthropic (May and July), Microsoft (July) and AWS (June) all launched deployment businesses during 2026. - Accenture completed its acquisition of Faculty, the London applied AI company, in March. - The ventures are aimed at the Fortune 500, FTSE 100 and private equity portfolios, with minimum engagements reported in the millions. - Model choice, knowledge boundaries and access are the three questions every buyer should now ask a deployment firm. - Outside the largest accounts, these moves strengthen the case for an independent applied AI firm. #### What happened Accenture completed its acquisition of Faculty, the London applied AI company, on 16 March 2026, bringing more than 400 AI-native professionals and the Frontier product into the consultancy. On 4 May, Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a joint venture of around $1.5 billion to build an AI services firm; it launched on 15 July as Ode with Anthropic. OpenAI followed on 11 May with the OpenAI Deployment Company, $4 billion of initial investment from 19 firms led by TPG, alongside Bain & Company, Capgemini and McKinsey, and an agreement to acquire Tomoro and its roughly 150 forward-deployed engineers. Amazon Web Services committed $1 billion to forward-deployed engineering. Days later, on 2 July, Microsoft launched the Microsoft Frontier Company with a $2.5 billion commitment and 6,000 industry and engineering experts. Five moves, and they point the same way. The organisations that make the models, and the organisations that host them, now deploy them inside customers as well. Sources for each are listed at the end of this article. #### Why the AI labs are doing it Margin, mostly. As models become interchangeable, the durable revenue sits in the services and operations around them, and the announcements say so openly. There is also a feedback loop: engineers inside customers see where the models fail, which improves the next model. And there is distribution. The investment firms behind each venture own thousands of portfolio companies and can steer them towards the venture, which is why private equity sits at the centre of both AI lab ventures. #### What it means for an enterprise buyer ##### Model choice is no longer independent of the firm deploying it A deployment company owned by an AI lab deploys that lab's models. Often that will be the right model anyway. But the choice has stopped being an evaluated one, and the model that wins on document extraction is rarely the one that wins on voice, or on code, or on long-context reasoning. The ranking changes every quarter. Ask any prospective firm how models get chosen, and ask to see the evaluation. ##### Knowledge boundaries need to be written down Forward-deployed engineers learn how your operation actually works. Where does that knowledge go when they leave? Reporting on the talent market has already noted that enterprises increasingly prefer internal teams to protect proprietary processes, and worry that external AI firms could use client knowledge competitively. The remedy is contractual: who owns the code, the data and the evaluation assets, and whether process knowledge may be reused across clients. ##### Access is rationed by size The ventures are built for the largest accounts. Industry coverage has reported minimum engagements in the millions of dollars, and every announcement names the Fortune 500, the FTSE 100 or private equity portfolios. If you are a mid-sized enterprise with a claims queue and a month-end close, none of this capacity is coming to you. #### What it means for mid-sized enterprises Most of the economy is not the Fortune 500. Mid-sized enterprises run the same kinds of process, answer to the same regulators and carry the same cost of getting AI wrong, and they are the organisations none of these ventures were designed around. They also have an advantage the largest accounts do not: a sponsor who can decide, a working group that can be assembled in a week, and a small enough estate that one process can be changed and measured this year. What they lack is engineering capacity on the frontier, which is faster to buy in than to build. #### How to choose a deployment firm now 1. **Model choice.** Who chooses the model, on what evidence, and can it be changed without a rebuild? 2. **Ownership.** Who owns the code, data, prompts and evaluation assets when the engagement ends? 3. **Knowledge boundaries.** Whether the firm may reuse your process knowledge with other clients. Get the answer into the contract. 4. **Independence.** Who owns the firm, and what else it sells you besides the engagement. 5. **Fit.** Is the engagement sized for your organisation, and does the firm serve companies like yours? 6. **After go-live.** Who runs the system, on what service levels, and how model upgrades get decided. #### Where Applicat AI stands Applicat AI is an independent frontier applied AI company, founder-owned, with no AI lab or cloud provider behind it and no model licence of its own to sell. We build on frontier models from Anthropic, OpenAI and Google and on open-weight models, chosen by evaluation on your cases. You own the assets, and we commit not to reuse your process knowledge with anyone else. We build for mid-sized enterprises, and we engage them directly. The full position is on the [Why independent](/why-independent) page. > **The short version.** The labs becoming services companies adds real capacity for the largest enterprises. For everyone else it is a reason to insist on independence: model choice made by evaluation, knowledge boundaries written into the contract, and a firm sized for your organisation. #### Sources - Accenture: Accenture Completes Acquisition of Faculty, 16 March 2026: https://newsroom.accenture.com/news/2026/accenture-completes-acquisition-of-faculty - TechCrunch: Anthropic and OpenAI are both launching joint ventures for enterprise AI services, 4 May 2026: https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/ - OpenAI: OpenAI launches the OpenAI Deployment Company, 11 May 2026: https://openai.com/index/openai-launches-the-deployment-company/ - TechCrunch: Microsoft launches its own AI deployment company with $2.5 billion commitment, 2 July 2026: https://techcrunch.com/2026/07/02/microsoft-launches-its-own-ai-deployment-company-with-2-5-billion-commitment/ - Business Wire: Anthropic, Blackstone, and Hellman & Friedman Introduce Ode with Anthropic, 15 July 2026: https://www.businesswire.com/news/home/20260715205134/en/Anthropic-Blackstone-and-Hellman-Friedman-Introduce-Ode-with-Anthropic-an-Enterprise-AI-Services-Firm - TechCrunch: Forward-deployed engineers are the AI industry's latest talent obsession, 30 July 2026: https://techcrunch.com/2026/07/30/forward-deployed-engineers-are-the-ai-industrys-latest-talent-obsession/ - The Next Web: OpenAI just acquired the consulting firm it was born alongside: https://thenextweb.com/news/tomoro-openai-deployment-company-consulting #### Questions on this article **What is the OpenAI Deployment Company?** A majority OpenAI-owned business launched on 11 May 2026 with $4 billion of initial investment from 19 firms led by TPG, which embeds forward-deployed engineers in organisations to build production systems on OpenAI models. Its founding acquisition was Tomoro, an applied AI consultancy. **What is Ode with Anthropic?** An enterprise AI services firm formed by Anthropic, Blackstone and Hellman & Friedman with Goldman Sachs and other investors, announced in May 2026 and launched on 15 July 2026, that combines Anthropic's models with embedded engineering teams for mid-sized organisations. **Should we still hire an independent applied AI firm?** If you want the model chosen by evaluation, your process knowledge kept under contract, and an engagement sized for your organisation, yes. Some organisations will use an AI lab's venture and an independent firm together, with the independent firm running the evaluation and the system after go-live. ### What is applied AI? A practical definition for 2026 (https://applicat.ai/insights/what-is-applied-ai) By Luka Kokot, Founder and Chief Executive, Applicat AI. Published 2026-09-02, updated 2026-09-15. Applied AI is the practice of taking existing AI capabilities, in particular frontier models, and applying them to specific business processes to produce measurable results in production. AI research creates new capabilities; AI strategy advises on where to use them; software integration builds to a specification. Applied AI is the one that has to answer for how a probabilistic system behaves once it is running inside real operations. Key takeaways: - Applied AI is measured by systems running in production with measured outcomes. A prototype does not count. - It combines four disciplines: use-case framing, engineering on frontier models, evaluation, and operations. - The most reliable indicator of an applied AI provider is who is accountable after go-live. #### A working definition **Applied AI** is the practice of taking existing AI capabilities, in particular frontier models, and applying them to specific business processes to produce measurable results in production. Three parts of that carry the weight. Existing capabilities: applied AI uses the best models available and sets out to create none of its own. The process has to be specific, which means a claims queue, a month-end close, a contact centre, and never "the business". And the results have to be measurable in production, because a system that only exists as a demonstration has not been applied to anything yet. #### How applied AI differs from research, strategy and integration Applied AI borrows from research, from strategy and from software integration, and is separated from all three by one thing: accountability for a probabilistic system running inside real operations. That is a heavier commitment than it sounds. It forces evaluation on real cases, human checkpoints where the risk sits, and an operating model for the years after go-live. | Discipline | Produces | Accountable for | | --- | --- | --- | | AI research | New models and methods | Capability | | AI strategy | Recommendations and roadmaps | Advice | | Software integration | Systems built to specification | Conformance to spec | | Applied AI | Production systems with measured outcomes | Results in production | #### Why the term matters now Frontier models became capable enough to do real work well before most organisations were able to put them to work. MIT's 2025 study of enterprise generative AI found that around 95% of pilots produced no measurable impact on profit and loss, while the small group that succeeded had chosen narrow processes, integrated deeply and worked with specialists. The gap is not a capability gap. It is an application gap, and applied AI is the discipline that closes it. #### The four disciplines inside applied AI 1. **Framing.** Selecting use cases by value, feasibility and data readiness, and writing a value case with a baseline and a target. 2. **Engineering on frontier models.** Designing the model layer, retrieval, tools, orchestration and integrations so the system works inside real constraints and survives model upgrades. 3. **Evaluation.** Building a test set from the organisation's own cases and scoring accuracy, safety and cost before go-live and on every change. 4. **Operations.** Monitoring quality and cost, managing model upgrades, responding to incidents and reporting value, month after month. #### What "frontier" adds to the definition A frontier applied AI company builds on the most capable models available, whichever lab produces them, and designs for the fact that the frontier moves. In practice that means model-agnostic architecture, model selection driven by evaluation, and a plan for upgrades. There is a second-order effect. The capability ceiling of a system rises over time without a rebuild, which is a different economic proposition from software that depreciates. #### How to recognise real applied AI - The proof runs on your data, inside your security perimeter. - There is an evaluation suite, and it is used to make decisions. - People are designed into the workflow at specific checkpoints, and those checkpoints are measured. - Someone is accountable for the system after go-live, with service levels and value reporting. - The architecture can change models without a rewrite. #### Sources - MIT NANDA, The GenAI Divide: State of AI in Business 2025: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf #### Questions on this article **Is applied AI the same as generative AI?** No. Generative AI is a class of capability. Applied AI is the practice of putting capabilities, including generative and frontier models, to work in specific processes with measured results. **Does applied AI require building our own models?** Rarely. Most applied AI systems use frontier models through enterprise endpoints, with retrieval, tools and evaluation around them. Fine-tuning is used selectively when evaluation shows a benefit. **What is a frontier applied AI company?** A company that builds applied AI systems on frontier models and takes responsibility for those systems in production. Applicat AI uses the term to describe itself. ### Why 95% of enterprise AI pilots never reach production, and what the 5% do differently (https://applicat.ai/insights/why-ai-pilots-fail-to-reach-production) By Luka Kokot, Founder and Chief Executive, Applicat AI. Published 2026-09-05, updated 2026-09-15. Most enterprise AI pilots fail for reasons that have nothing to do with the models. They are built outside production constraints, they have no evaluation, nobody owns the outcome, and the organisation around them never changes. The ones that succeed pick narrow processes, prove on real data, work with specialists, and run the system as an operation for as long as it exists. Key takeaways: - MIT NANDA (2025): about 95% of enterprise generative AI pilots show no measurable P&L impact. - The same study found deployments built with external specialists succeed roughly twice as often as internal builds. - Gartner (2024) forecast that at least 30% of generative AI projects would be abandoned after proof of concept by end-2025. - The fix is a method that proves in production conditions, evaluates on real cases and assigns ownership after go-live. #### The numbers In August 2025, MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025. The headline finding: around 95% of enterprise generative AI pilots delivered no measurable impact on profit and loss, despite tens of billions of dollars of enterprise investment. A year earlier, Gartner had predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. Both numbers describe one phenomenon from different angles. Generative AI is easy to start and hard to finish. The MIT study also carried the most useful detail for anyone trying to be in the 5%: deployments built with specialised external firms succeeded about twice as often as internal builds, and the successful group concentrated on narrow, high-value processes and integrated deeply into workflows. #### Why pilots fail: four structural causes ##### 1. The pilot was built outside production conditions Sample data, a sandbox, no identity integration, no audit trail, no cost profile. The demonstration works, and then every one of those omissions becomes a project of its own. By the time integration, permissions and security review are complete, the sponsor has moved on. This is the pilot trap. It is the single most common failure mode we see. ##### 2. There was no evaluation Without an evaluation suite built from the organisation's own cases, nobody can say whether the system is good enough, whether it got worse after a change, or whether a different model would do better. Decisions revert to opinion, and opinion does not survive a risk committee. ##### 3. Nobody owned the outcome A strategy firm framed it, a vendor supplied a licence, an integrator built to specification and an internal team inherited it. When performance fell short, each could point at another. Systems whose behaviour is probabilistic need a single accountable owner with the authority to change the model, the prompts, the tools and the process. ##### 4. The organisation was not changed MIT's study noted that the tools that stalled could not retain feedback, adapt to context or improve over time, and that adoption of generic tools was high while transformation was rare. A system that is dropped into an unchanged process, with no checkpoint design, no training and no adoption measurement, is optional. Optional systems are not used. #### What the 5% do differently - **They choose narrow processes with a number attached.** A claims queue, a reconciliation, a category of customer request, with a baseline someone already reports on. - **They prove on real data inside the perimeter.** The proof is the first release, so there is no second project to reach production. - **They evaluate before, during and after.** Every decision about go-live, model choice and scope is made on measured results. - **They design people in.** Checkpoints are defined, fast and measured. Judgment stays with people; throughput moves to the system. - **They work with specialists.** Twice the success rate in MIT's sample. The frontier moves too quickly for most organisations to keep a bench current. - **They run it as an operation.** Monitoring, regression evaluation, model upgrades and monthly value reporting, for as long as the system exists. #### A method that avoids the trap The Applicat Method encodes those six behaviours into five stages: Frame, Prove, Build, Deploy and Run. Each stage has a defined output and ends with a decision made on evidence. The Prove stage in particular is built to be the first production release, which is why proofs convert. You can read the full method on the [method page](/method). > **The test to apply to any AI proposal.** Ask three questions. Where does this run, on what data, and who is accountable after go-live? If the answers come back as a sandbox, a sample and nobody, you are looking at a pilot. #### Sources - MIT NANDA, The GenAI Divide: State of AI in Business 2025: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf - Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025", July 2024: https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025 #### Questions on this article **What did the MIT GenAI Divide report actually measure?** It examined enterprise generative AI adoption through interviews, surveys and analysis of deployments, and found that around 95% of pilots produced no measurable P&L impact. It attributed success to narrow process selection, deep integration and working with specialised vendors. Model quality was not the differentiator. **Is a 95% failure rate normal for new technology?** High early failure rates are common with transformational technology, but the MIT and Gartner findings point to fixable causes: proofs built outside production conditions, no evaluation, no ownership and no organisational change. Every one of those is a method problem. **How do we know if our pilot is at risk?** If it runs on sample data outside your security perimeter, has no evaluation suite, and no named owner for the period after go-live, it is at risk. Each of those can be fixed before more money is spent. ### Frontier models in the enterprise: how to choose, evaluate and switch without rebuilding (https://applicat.ai/insights/choosing-frontier-models-for-the-enterprise) By Luka Kokot, Founder and Chief Executive, Applicat AI. Published 2026-09-10, updated 2026-09-15. Enterprises should not pick a frontier model the way they pick a database vendor. Models change every few months. What works instead is selection driven by evaluation on your own cases, an architecture that isolates the model behind tested interfaces, and an operating rhythm in which upgrades are regression-tested and rolled out like any other change. Key takeaways: - Choose models by running them against your own cases. Public leaderboards rank general ability, which is a different question. - Isolate the model behind an interface. Then a swap is a test run. - Plan for upgrades every quarter. The frontier moves that fast. - Keep at least one open-weight model evaluated and ready. - Open weights earn their place on data residency, on latency, and on volumes where per-token pricing stops working. #### The problem with choosing a model Every few months an AI lab releases a model that is better, cheaper or faster than the one before, and often all three. Public benchmarks measure general ability. They say nothing about your claims letters or your product catalogue. An enterprise that standardises on one model in January and hard-wires it into ten systems will spend the following year either falling behind or rebuilding. #### Principle one: evaluate on your own cases Build an evaluation set from real cases: documents, requests, questions and expected outcomes drawn from your operation, with the awkward ones included. Score candidate models on accuracy, groundedness, safety behaviour, latency and cost at expected volumes. The result is a ranking that means something for your business, and it is reusable every time a new model arrives. #### Principle two: isolate the model Treat the model as a component behind an interface. Prompts, tool definitions, retrieval and output schemas live in your system, versioned and tested. The model is called through a gateway that can route to Anthropic, OpenAI, Google or an open-weight model running in your tenancy. Swapping models then becomes a configuration change, validated by the evaluation suite before it goes anywhere near production. #### Principle three: run upgrades as operations 1. A new model is released. 2. It is run against the evaluation suite in a staging environment. 3. Results are compared with the current model on accuracy, safety, latency and cost. 4. If it wins, it is rolled out to a share of traffic, monitored, then promoted. 5. The change is recorded in the system's governance log. This is the same discipline that infrastructure teams apply to patches. It turns the fast-moving frontier from a risk into a source of routine improvement: the same system gets better every quarter without a rebuild. #### Where open-weight models fit Open-weight models run inside your own environment, which matters when data cannot leave a jurisdiction, when latency must be very low, or when volumes make per-token pricing uncompetitive. They are usually behind the frontier on general capability but can match or beat it on narrow tasks after fine-tuning. A mature enterprise portfolio keeps at least one open-weight option evaluated and ready. #### Questions to ask any AI provider - Which models did you evaluate for this use case, on what cases, and can we see the scores? - How long would it take to switch models, and what would break? - What is your process when a new model is released? - Where does our data go for each model you propose? - What does inference cost at our expected volumes, and how is that monitored? > **The Applicat AI position.** We are model-agnostic by design. We build on the frontier, select by evaluation on your cases, and treat model upgrades as a routine part of Managed AI Operations. #### Questions on this article **Which frontier model is best for enterprise use?** There is no single answer. The best model for a use case depends on the task, the data, the latency and cost targets and the residency rules. Evaluate candidates on your own cases and re-evaluate when new models are released. **Should enterprises use more than one model provider?** Usually yes. A model-agnostic architecture with a gateway lets you route each use case to the model that evaluates best and reduces dependency on any one provider. **How often should we re-evaluate models?** At least quarterly, and whenever a major release arrives. With an evaluation suite already in place, a re-evaluation takes hours. ### Agentic AI governance: a ten-point checklist before an agent goes into production (https://applicat.ai/insights/agentic-ai-governance-checklist) By Luka Kokot, Founder and Chief Executive, Applicat AI. Published 2026-09-12, updated 2026-09-15. AI agents that take actions need a different governance model from chat assistants that produce text. This checklist sets out the ten controls Applicat AI puts in place before any agent goes live: they cover access, tools, human checkpoints, evaluation, logging, cost, incidents, change control, data handling and ownership. Key takeaways: - An agent that acts needs the same controls as a person with system access, plus evaluation. - Design human checkpoints around consequence. An agent that drafts a summary and an agent that releases a payment do not need the same gate. - Log every action, make it attributable and, where possible, reversible, in a record you can put in front of an auditor. #### Why agents change the governance question A chat assistant produces text that a person then acts on. An agent acts: it creates the ticket, updates the record, sends the email or approves the request. That moves the governance question from "is the answer accurate" to "is the action authorised, bounded, reviewable and reversible". The controls below are the ones we require before an agent enters production, whatever the use case. #### The ten controls 1. **Least-privilege identity.** The agent has its own identity with the minimum permissions for its task, reviewed like any service account. 2. **Bounded tools.** The agent can only call an explicit list of tools, each with defined inputs, limits and failure behaviour. 3. **Human checkpoints by consequence.** Actions are classified by impact and reversibility; consequential or irreversible actions require a person to approve, with the context needed to decide quickly. 4. **Evaluation gates.** No change to a model, prompt, tool or retrieval source reaches production without passing the evaluation suite. 5. **Complete logging.** Every step, tool call, input and output is logged with a trace identifier, retained according to policy and searchable. 6. **Cost and rate limits.** Per-task and per-day budgets and rate limits prevent runaway loops and surprise invoices. 7. **Incident response.** A defined path to pause the agent, roll back actions where possible, notify owners and review, with service levels. 8. **Model change control.** Model upgrades follow a documented process: staging evaluation, comparison, phased rollout, record. 9. **Data handling rules.** What data the agent may read, what it may send to which model, where it is processed and how long it is retained, all documented and enforced. 10. **A named owner.** One accountable person for the agent's behaviour, its metrics and its changes, with the authority to stop it. #### What the record should look like Governance that lives in a policy document is not governance. For each agent in production there should be a living record: its purpose and value case, its permissions and tools, its checkpoint design, its evaluation results over time, its change log and its incidents. That record answers internal audit, customer due diligence and emerging regulation without a scramble. #### A note on proportionality The controls scale with consequence. An agent that drafts internal summaries needs lighter checkpoints than one that releases payments. The mistake is not too little governance or too much; it is governance that is uniform, which makes low-risk agents slow and high-risk agents no safer. Classify actions by impact, then design the checkpoints to match. > **How Applicat AI applies this.** These ten controls are built during the Build stage of the Applicat Method and monitored during the Run stage. Security and risk teams get a finished package to review. #### Questions on this article **Do AI agents need human approval for every action?** No. Approval should be required for consequential or irreversible actions, and designed to be fast. Low-impact, reversible actions can be automated and reviewed through logs and sampling. **How do you stop an AI agent from running up costs?** With per-task and per-day budgets, rate limits, loop detection and alerts, all enforced in the orchestration layer. **Who should own an AI agent in production?** A named business or operations owner accountable for its outcomes, supported by an operations team that monitors quality, cost and changes. Ownership must include the authority to pause the agent. ## About Applicat AI (https://applicat.ai/about) Applicat AI is a frontier applied AI company, founded in London in 2026 and owned by the people who run it. We put frontier models to work inside mid-sized enterprises one process at a time, on the client's own systems, and we stay with each system after it goes live. The company is young. This page sets out who runs it, how an engagement is staffed and what we commit to. ### The real economy runs on documents, requests and decisions. Banks, operators, manufacturers, hotel groups, professional firms and public bodies do their work in paperwork, queues and judgment calls. Our mission is to have frontier models doing a real share of that work in production, measured against a number the business already tracks, and getting better while they run. ### Story Applicat AI was founded by Luka and Marko Kokot with the team behind BroadVision Technologies, a global IT services company with more than 26 years of enterprise delivery across the United Kingdom, Europe and Africa. That heritage matters for one reason. We have spent decades being answerable for systems that businesses depend on every day, and AI has not changed what being answerable means. The company started from a pattern we kept seeing. Frontier models had become capable enough to do real work, and organisations were still stuck between a demonstration that impressed everyone and a production system that never arrived. The missing piece was never a better model. It was a team that would work out which use case was worth doing, prove it on real data, build it to the standards your security review applies to everything else, and then own it once it was live. We are independent by design. The founders hold the company, and no AI lab, cloud provider, consultancy or investment fund holds a share of it. In the year the labs and the clouds became deployment companies themselves, that is what lets us choose a model by testing it, and keep what we learn about a client inside that client. We build for mid-sized enterprises and we go to them directly. Their operations are large enough for an AI system to move a number leadership already watches, and they can still decide in a room instead of across a year of committees. The deployment arms of the AI labs and the global consultancies were not built to serve them, which is the gap we were built for. And we are new. Applicat AI was founded in 2026, which is a fact a buyer should weigh, so we put it near the top of this page. What a young company can offer in place of a logo wall is everything about itself that can be checked today, and an introduction to someone who has worked with us. ### Principles - **Production over demos.** A system that is not in production has created nothing. We measure ourselves on what runs. - **Evidence over opinion.** Model choices, go-live decisions and scope changes are settled by results from tests built out of your own cases. - **Judgment stays human.** Agents take throughput and synthesis. People keep accountability, taste and the decisions that carry consequence. - **Build on the frontier, independently.** We use the most capable models any lab has, chosen by testing, and design the system so that next year's model is an upgrade you can prove in an afternoon. - **Own the outcome.** The team that frames the work builds it and runs it. There is no hand-off where responsibility goes missing. - **Sized for the organisations we serve.** Mid-sized enterprises get the engineering depth usually reserved for the largest accounts, in engagements scoped to their operations and their budget. ### Leadership Luka Kokot, Co-founder and Chief Executive. Technology executive and entrepreneur. Electrical engineer by training, with a Master's degree in AI and management. Luka leads global IT services and platform businesses across the United Kingdom, Europe and Africa, and founded Applicat AI to close the gap between what frontier models can do and what is actually in production inside organisations. Marko Kokot, Co-founder and Chief Technology Officer. Electrical engineer, trained at Queen Mary University of London and currently reading for a Master's in Applied Machine Learning at Imperial College London. Marko oversees technology at Applicat AI: how systems are built, which models earn their place in them, and what has to be true before anything reaches a client's users. ### Four named roles, one team that stays. You meet the people who will do the work before you sign, and the same team carries it from the first workshop to the monthly report. - **Engagement lead.** A senior practitioner who owns the outcome end to end, runs the decision points with your sponsor, and has the standing to put a recommendation against a use case in writing. - **Forward-deployed engineers.** Applied AI engineers who work inside your teams and your tools to frame, prove and build the system, and who hand what they build to your own people as they go. - **Solution architect.** Accountable for how the thing is put together: the model layer, the integrations, who is allowed to see what, the safety limits, and the tests every change has to pass. - **Operations engineer.** Joins at Deploy and stays for Run: watching quality and cost, re-testing after every change, managing model upgrades and answering the phone when something breaks. ### Careers We hire engineers who want their name on something that runs. We are a distributed team with headquarters in London. Roles open at the moment: - Forward-deployed applied AI engineer (London or remote (UK, Europe, Africa)) - Applied AI solutions architect (London or remote) - Applied AI operations engineer (Remote) - Applied AI engagement lead (London or remote (UK, Europe, Africa)) ### Questions about Applicat AI **When was Applicat AI founded?** Applicat AI was founded in 2026 in London, United Kingdom, by Luka and Marko Kokot with the team behind BroadVision Technologies. It is a young company and says so: named case studies will be published as engagements complete and clients agree to be named. **Is Applicat AI part of BroadVision Technologies?** Applicat AI is an independent company. It was founded by the team behind BroadVision Technologies, a global IT services company, and draws on that team's experience of enterprise delivery, but it operates as its own business, with its own clients and its own delivery team. **How large is the Applicat AI team?** We are a growing, globally distributed team of applied AI engineers, solution architects and operations engineers, headquartered in London. Every engagement is staffed with named engineers, and you meet them before you sign anything. **Is Applicat AI owned by an AI lab, a cloud provider or a consultancy?** No. Luka Kokot and Marko Kokot own the company between them, and nobody else holds a share of it. That is why the model used on an engagement is decided by testing candidates on the client's own cases, and why a piece of work can end with us recommending that the client does not build. If the ownership ever changes, the why independent page will say so before we say it anywhere else. ## Frequently asked questions (https://applicat.ai/faq) Direct answers to the questions mid-sized enterprises ask us most often: what we do, how an engagement runs, what it costs, where your data goes, which models we build on and who owns the company. If your question is not here, contact us and we will answer it personally. ### About Applicat AI **What is Applicat AI?** Applicat AI is a frontier applied AI company headquartered in London. We design, build and run production AI systems on frontier models for mid-sized enterprises, and we stay accountable for those systems after go-live. Most engagements begin with one process that is costing an organisation real money and end with a working system inside that organisation, reported on every month against the number agreed at the start. **What does "applied AI" mean?** Applied AI is the practice of taking AI capabilities that already exist, in particular frontier models, and applying them to specific business processes to produce measurable results in production. It is distinct from AI research, which creates new capabilities, and from AI strategy, which advises on them. Success is measured by what is running in your business at the end of it. **What does "frontier" mean in "frontier applied AI company"?** Frontier models are the most capable AI models available at any given time, currently produced by AI labs such as Anthropic, OpenAI and Google, alongside the strongest open models. A frontier applied AI company builds on those models and designs its systems so that the next generation can be swapped in and proven in days, which matters because the frontier moves every few months. **Who owns Applicat AI?** The founders, and no outside shareholder of any kind: in particular none of the model labs, clouds, consultancies or funds now selling AI deployment. We also sell no software licences of our own. That is why the model that goes into your system is the one that scored best on your own cases, and why we are free to tell you that a thing is not worth building. **Why choose an independent applied AI firm over an AI lab's deployment company?** Because the model is then picked by testing it on your work, your process knowledge and everything built with it stay yours under contract, and the engagement is sized for your organisation rather than for the Fortune 500. Plenty of organisations will use both: an independent firm can run the testing, and run the system after go-live, alongside a lab's own team. **How is Applicat AI different from an AI consultancy?** A consultancy advises. Applicat AI frames the work, proves it, builds it, deploys it and then runs it, with one team accountable from the first workshop to the monthly report in production. ### Working with us **How does an engagement start?** With a conversation about the outcome you need, followed by an Applied AI Sprint: four to six weeks, fixed scope, fixed fee. It produces a ranked shortlist of what is worth building and a working version of the strongest candidate, running on your real data inside your own environment. **How long until we have something in production?** An Applied AI Sprint puts a working, production-grade system on your real data within four to six weeks. What follows is scoped once that has been measured, so the date for a first release to your users is agreed on evidence once there is some. It depends mostly on your integrations and your governance requirements. **What does it cost?** Sprints are a fixed fee. Programmes are priced against outcome milestones. Managed AI Operations is a monthly subscription with service levels attached. We give indicative ranges in the first conversation, before anyone spends time on a proposal. **Do you work with organisations outside the United Kingdom?** Yes. Our engineering team is globally distributed and we work with clients across the UK, Europe, Africa, the Middle East and North America. **Can our own engineers be involved?** We encourage it. Our forward-deployed engineers work alongside your teams, and handing over what we build is part of the Build and Deploy stages. You decide afterwards how much of the running you keep in-house and how much you leave with us. **What happens if we end the relationship?** The system keeps running. It sits in your own environment, on your accounts, and the code, the data, the prompts and the test cases are yours in the contract, with your team holding what it needs to change them. There is no part of it we can switch off. ### Technology, security and data **Which AI models does Applicat AI use?** We have no house model. We build on frontier models from Anthropic, OpenAI and Google, and on the leading open models when data residency, cost or speed favour them. Which one goes into your system is settled by running your own cases through the candidates and comparing the scores, and it can change when a better model arrives. **Where does our data go?** Nowhere it is not already. The system runs in your own cloud account, or in a dedicated environment under your control, and model access uses enterprise endpoints where your data is not used to train anyone's models. Residency rules are captured in the first week and drive the choice of region and model. **How do you make AI systems safe to use in production?** Every system gets the least access it needs, a fixed set of tools it is allowed to use, checks on what it sends back, a person in the loop wherever an action is hard to undo, and a full record of what it did and why. Every change is then re-run against tests built from your own cases, which is the evaluation suite that gates releases. These are built during Build and watched during Run. **Do you integrate with our existing systems?** Yes, and the work usually stands or falls on it. We integrate with Microsoft 365 and Azure, Google Workspace and Cloud, AWS, SAP, Microsoft Dynamics, Sage, Salesforce, ServiceNow, HubSpot, Zendesk and most platforms with a supported API, and with secure automation where an API does not exist. ## Applied AI glossary (https://applicat.ai/glossary) Short, precise definitions of the terms that appear in applied AI work, written for business leaders and engineers who need a shared vocabulary. Each definition is the one Applicat AI uses in its own engagements. - **Applied AI.** The practice of taking existing AI capabilities, in particular frontier models, and applying them to specific business processes to produce measurable results in production. Applied AI sits between research, which creates new capabilities, and strategy, which advises on them. Its unit of success is a system running in production with measured outcomes. - **Frontier model.** One of the most capable AI models available at a given time, typically produced by a small number of AI labs such as Anthropic, OpenAI and Google, alongside the strongest open-weight models. Because the frontier moves every few months, systems built on frontier models should isolate the model behind evaluated interfaces so that upgrades are tests rather than rebuilds. - **Frontier applied AI company.** A company that builds applied AI systems on frontier models and takes responsibility for those systems in production. Applicat AI describes itself with this term: frontier because of the models it builds on, applied because its work is measured by systems in production. - **AI agent.** A software system that uses a model to decide on and carry out actions towards a goal, calling tools such as search, databases and application interfaces to do so. What distinguishes an agent is that it takes actions in systems of record. - **Agentic workflow.** A multi-step business process in which AI agents read inputs, decide next actions and execute them through integrations, with people at defined checkpoints. - **Human-in-the-loop.** A design in which a person reviews, approves or corrects an AI system's output or action at defined points, typically where the consequence of an error is high. A well-designed checkpoint gives the reviewer the evidence and the time to make a real decision in seconds. - **Evaluation (evals).** A repeatable set of test cases and scoring methods used to measure an AI system's accuracy, safety, cost and behaviour, run before go-live and on every change. Evaluation suites built from an organisation's own cases are what turn model choice and go-live decisions from opinion into evidence. - **Guardrails.** Controls that constrain what an AI system may read, say or do: input filtering, output checks, tool permissions, rate and cost limits and escalation rules. - **Retrieval-augmented generation (RAG).** A technique in which relevant passages are retrieved from an organisation's own content at question time and supplied to the model as context, so answers are grounded in those sources and can be cited. - **Groundedness.** The degree to which an AI system's answer is supported by the sources it was given, as opposed to invented. Measured as part of evaluation. - **Hallucination.** An output that is fluent and confident but not supported by the source material or by fact. Reduced through grounding, citations, refusal behaviour and evaluation. - **Embeddings.** Numerical representations of text, images or other content that place similar items close together, enabling semantic search and retrieval. - **Vector database.** A data store optimised for searching embeddings by similarity, used to retrieve relevant content for AI systems. - **Tool use.** The ability of a model to call functions, APIs or applications, for example to look up an order, create a ticket or run a query, as part of producing a result. Also called function calling. - **Model Context Protocol (MCP).** An open standard for connecting AI models and agents to tools and data sources through a common interface, reducing bespoke integration work. - **Orchestration.** The layer that coordinates models, tools, retrieval, memory and checkpoints into a working system, including error handling, retries and logging. - **Observability.** The instrumentation that makes an AI system's behaviour visible in production: traces of each step, quality scores, latency, cost and failure modes. - **Model-agnostic.** An architecture and delivery approach that is not tied to one model provider, choosing and switching models on the basis of evaluation. - **Open-weight model.** A model whose weights are published, so that it can be run inside an organisation's own environment. Often called simply an open model, and usually chosen for data residency, cost or latency reasons. - **Fine-tuning.** Further training of an existing model on specific examples to specialise its behaviour. Used selectively in applied AI, when evaluation shows it outperforms prompting and retrieval. - **Context window.** The amount of text, measured in tokens, that a model can consider at once, including instructions, retrieved content and conversation history. - **Token.** The unit in which models read and produce text, roughly three-quarters of an English word. Model usage and cost are measured in tokens. - **Inference.** Running a trained model to produce an output. Inference cost and latency are core operating metrics of an AI system in production. - **Multimodal.** A model or system that works with more than one kind of input, such as text, images, documents, audio and video. - **Forward-deployed engineer.** An engineer who works inside a client organisation, in its tools and its meetings, designing and building systems against the constraints as they actually are on the ground. - **Managed AI operations.** The ongoing operation of AI systems after go-live: quality and cost monitoring, regression evaluation, model upgrade management, incident response and value reporting. - **Business operating system (BOS).** The layer a company builds and runs its own AI applications and agents on: governed connections to its systems, one identity and permissions model, permission-aware retrieval, one model gateway, a build surface, an evaluation harness, an audit trail and a register of what exists. Applicat AI builds a business operating system inside a client's own cloud tenancy, so that the things every AI application needs exist once, under the client's control, and each new application inherits them. The same phrase is also used for a management framework of meetings and metrics, which is a different thing. - **Regression testing.** Re-running an evaluation suite after any change to a model, prompt, retrieval source or tool to confirm that behaviour has not degraded. - **AI governance.** The policies, controls and records that determine how AI systems are approved, monitored and changed, including accountability, human oversight and documentation for regulators and auditors. - **Data residency.** Requirements about the geographic location in which data may be stored and processed, which shape model, region and architecture choices for AI systems. - **Pilot trap.** The pattern in which AI prototypes succeed as demonstrations but never reach production because they were built outside real data, integration, security and governance constraints. The Applicat Method avoids the pilot trap by proving on real data inside the production perimeter from the start. - **Containment rate.** In customer and service operations, the share of requests resolved by an AI system without a person becoming involved. - **Value case.** A statement of the measurable business outcome an AI use case is expected to change, the baseline, the target and how it will be measured in production. - **Digital coworker.** An AI agent that works alongside people on real tasks inside an organisation's systems, taking throughput and synthesis while people keep judgment and accountability. ## Contact (https://applicat.ai/contact) Tell us about the process you want to change and the number you want it to move. Every enquiry reaches a person who can answer it, and we reply within one business day. Enquiries: info@bivi.tech. Headquarters: London, United Kingdom. LinkedIn: https://www.linkedin.com/company/applicat-ai. We reply within one business day. Reasons people get in touch: - I lead or advise an organisation and want to discuss an applied AI engagement - I am interested in working at Applicat AI - Something else ## Privacy policy (https://applicat.ai/privacy) How Applicat AI collects, uses and protects personal data on this website and in its business relationships, in line with UK GDPR. Last updated 15 September 2026. This privacy policy explains how Applicat AI Ltd ("Applicat AI", "we", "us") collects and uses personal data when you visit https://applicat.ai, contact us, or work with us as a client or supplier. We are the data controller for the personal data described here. ### What we collect - Contact details you give us, such as your name, work email address, company and role, when you submit a form or email us. - Correspondence and the content of enquiries, proposals and meetings. - Technical data such as IP address, browser type and pages visited, collected through server logs and, where enabled, privacy-respecting analytics. ### How we use it - To respond to enquiries and manage client and supplier relationships (legitimate interests and performance of a contract). - To improve the website and understand how it is used (legitimate interests). - To send occasional information about our services where you have asked for it or where we have a legitimate interest and you can opt out at any time. - To comply with legal obligations. ### Sharing and processors We share personal data only with service providers that act on our instructions, such as hosting, email and customer relationship tools, and with professional advisers where necessary. We do not sell personal data. Where data is transferred outside the United Kingdom we use appropriate safeguards such as the UK International Data Transfer Agreement or adequacy regulations. ### Retention We keep enquiry data for up to 24 months after our last contact, and client and supplier records for the duration of the relationship plus the period required by law or for the defence of legal claims. ### Your rights Under UK GDPR you have the right to access, correct, erase or restrict the processing of your personal data, to object to processing based on legitimate interests, and to data portability. To exercise these rights, email info@bivi.tech. You also have the right to complain to the Information Commissioner's Office (ico.org.uk). ### Cookies This website uses only strictly necessary cookies by default. If analytics or marketing cookies are introduced, a consent mechanism will be provided and this policy updated. ### Contact Questions about this policy can be sent to info@bivi.tech, or by post to Applicat AI Ltd, London, United Kingdom. ## Terms of use (https://applicat.ai/terms) Terms governing the use of the Applicat AI website. Last updated 15 September 2026. These terms govern your use of https://applicat.ai (the "Site"), operated by Applicat AI Ltd. By using the Site you agree to these terms. ### Use of the Site You may use the Site for lawful purposes only. You must not attempt to gain unauthorised access to the Site, interfere with its operation, or use automated means to extract content beyond what is permitted for search and AI indexing in our robots.txt file. ### Content Content on the Site is provided for general information about Applicat AI and applied AI. It does not constitute professional advice, and it does not form part of any contract. Statistics attributed to third parties are cited with their sources and remain the work of those parties. ### Intellectual property The Site and its content are owned by Applicat AI Ltd or its licensors. You may quote short extracts with attribution and a link to the source page. Any other reproduction requires our written permission. ### Liability The Site is provided as is. To the extent permitted by law, we exclude liability for loss arising from use of the Site or reliance on its content. Nothing in these terms limits liability that cannot be limited by law. ### Governing law These terms are governed by the laws of England and Wales, and the courts of England and Wales have exclusive jurisdiction. ### Contact Questions about these terms can be sent to info@bivi.tech.