# Governance works best when it is part of the system design.

> Put the company’s AI rules into everyday practice: who may use which information, what the software may do and how changes are checked and approved.

Canonical URL: https://marmation.com/expertise/ai-data-and-model-governance/

## The decision

Does the software enforce the rules your company says it follows?

For each important policy, identify who is responsible, which setting or code enforces it and how it will be tested. A log of an unauthorized action is evidence of a failure, not a working safeguard.

## Who approved what this assistant can read and change?

An AI assistant can read restricted records, draft suggestions and call APIs. The organization needs to know who authorized each capability and how a consequential action is reviewed.

## A policy and a log cannot prevent an unauthorized action

High-level responsible AI language can create false reassurance if production controls are weak. A system may log everything but still allow a user to see restricted information.

## Connect each policy to an enforced control and an owner

Agree which failures matter for this particular application, then map its data, model calls, tools, approvals and incident response. Link each written rule to a responsible person and an enforced control. Use relevant NIST or OWASP guidance to examine the design, while keeping compliance claims separate from the presence of a framework reference. Check that monitoring and change procedures reflect what the running system actually permits.

## Records and responsibilities that make oversight usable

- AI system inventory and risk/decision register.
- Data access, model/tool boundary and approval architecture.
- Evaluation, logging, change approval and incident requirements.
- Evidence package and operating responsibility mapping.


## Give consequential actions stronger controls

Governance should be proportional. A read-only knowledge assistant is not the same risk as an agent changing financial records. Avoid heavyweight controls that do not address a credible failure mode.

## Check that the running system follows the written rules

Test high-consequence actions, permission changes, injection, output misuse and audit completeness. Validate that operational practices match written control statements.

## Trace one action from request through approval and recovery

For each AI action, ask: who can initiate, what can it read, who can approve, what gets recorded and how can it be undone?

## Related reading

- [Give an agent a job. Give the system boundaries.](https://marmation.com/solutions/ai-agents-and-workflows/)
- [Make AI work with the systems your business already runs.](https://marmation.com/solutions/enterprise-ai-systems/)
- [Autonomy should be earned, not assumed.](https://marmation.com/expertise/ai-agent-orchestration/)