Enterprise AI reviews regularly raise three questions: where does the data go, who is accountable, and has the model accessed production records it should not have? Pilot teams struggle to answer when governance was omitted from the original build. Recording these decisions during development makes the review more useful.
A small set of project defaults can make governance easier to implement consistently. The controls below are the baseline we recommend to customers in regulated and public-sector work. Establishing them early reduces the rework needed when security, legal, and data teams review the deployment.
The eight defaults
1. Data-residency tags on every prompt
Give each inbound request a residency tag, such as US, EU, IN, or GOV. Configure the router to reject provider or region choices that do not match the approved policy. Test that behavior as part of deployment.
2. Provider allow-lists per workload
Approve providers per workload. A workload touching PII should not fall back to a model that has not been contractually approved. Keep the allow-list in configuration so the approved choices can be inspected and changed through review.
3. Output classification hooks
Before an output leaves the agent, run a classifier for PII leakage, policy violations, and jailbreaks. Version the checks, log their results, and evaluate their coverage as part of the release process.
4. Traces on every step, retained for 90 days
Send each agent step to a trace store, including the prompt, tool call, model response, latency, and cost. The proposed 90-day retention period provides a starting point for incident investigations; review it against the workload's privacy, retention, and audit requirements.
5. A human-in-the-loop escape hatch
Every production agent needs a path to a person when it cannot complete a task reliably. Make that handoff visible in the workflow and logs. It gives operators a way to review cases that the original design did not anticipate.
6. Documented accountable owner
Record a name, team, and escalation contact for each agent. The accountable owner should have authority to investigate quality issues and change or suspend the deployment.
7. Model-change notifications with a rollback plan
Providers deprecate and retune models. Subscribe to their change feeds, and document a rollback model for every agent. Use evaluations and production monitoring to identify regressions before they affect more customers.
8. A signed-off risk register
Maintain a one-page risk register for each agent: what it does, who it affects, what data it touches, what could go wrong, and which controls address those risks. Have the accountable owner sign it and review it quarterly so reviewers can see the current decisions.
What this does not cover
These eight defaults are a starting point. A mature AI governance program also covers training-data lineage, model-card management, red-team exercises, and third-party model assessments. Build those activities around the same ownership and review processes as the initial controls.
