AI governance is often sold as a 200-page framework. Teams of 3–10 people need something else: clear limits, evidence and owners.
The non-negotiable minimum
- Who can deploy an agent or change its tools.
- What data the model sees (and what it must never see).
- Which actions it can run without a human.
- Where the trail lives (prompt, tools, approval, outcome).
- Who can shut it down on a Friday at 6pm.
If you cannot answer that on one page, you do not have governance: you have hope.
EU AI Act without panic
Not everything is “high risk”. Many internal support or productivity uses fall under proportionate transparency and risk-management duties. The useful move is not memorising articles: it is documenting purpose, data, human oversight and an incident procedure.
One good internal page > a vendor PDF nobody reads.
Permissions > prompts
A prompt is not an access control. If the agent has write credentials to the ERP, the model can “decide” to use them. Prefer:
- Least-privilege service accounts
- Tool allowlists per environment
- Draft / execute separation
- Secrets outside the model context
Logs that help you learn
Keep enough to reproduce a failure: input, tools called, human decision, output. You do not need the whole embedding space. You do need to say: “on Tuesday the agent proposed X; Ana rejected it; the reason was Y”.
Small-team cadence
- Biweekly review of agent incidents (15 minutes).
- Living list of forbidden actions.
- A named owner (not “the team”).
- Documented kill switch.
Governance is not bureaucracy: it is making failure cheap to detect and reverse. For small teams, that is an edge over whoever ships and looks away.