Governance
Agentic AI governance: a ten-point checklist before an agent goes into production
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.
- Author
- Luka Kokot, Founder and Chief Executive, Applicat AI
- Published
- Updated
- Reading time
- 6 min read
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
- Least-privilege identity. The agent has its own identity with the minimum permissions for its task, reviewed like any service account.
- Bounded tools. The agent can only call an explicit list of tools, each with defined inputs, limits and failure behaviour.
- 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.
- Evaluation gates. No change to a model, prompt, tool or retrieval source reaches production without passing the evaluation suite.
- Complete logging. Every step, tool call, input and output is logged with a trace identifier, retained according to policy and searchable.
- Cost and rate limits. Per-task and per-day budgets and rate limits prevent runaway loops and surprise invoices.
- Incident response. A defined path to pause the agent, roll back actions where possible, notify owners and review, with service levels.
- Model change control. Model upgrades follow a documented process: staging evaluation, comparison, phased rollout, record.
- 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.
- 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.