AI is quickly becoming useful for the grunt work of data governance: classifying assets, suggesting owners, flagging policy violations. That's genuinely valuable — governance has always been under-resourced. But it raises a question regulated organizations can't hand-wave: when an agent makes a governance decision, who is accountable for it?
“The model did it” is not an answer
In defense, finance, and healthcare, every decision about sensitive data needs an owner. If an agent reclassifies a dataset or grants a role and no one can say why, that's not automation — it's an audit finding waiting to happen.
Treat agents as named actors
The fix is to stop treating AI as invisible background automation and start treating each agent as a first-class participant with a name, a defined scope of authority, and its own audit trail. Then an agent's decision is exactly as accountable as a human steward's: attributable, bounded, and logged.
Autonomy should be a dial, not a switch
Not every domain deserves the same level of independence. A useful model has three tiers you can set per domain:
- Advisory — the agent recommends, a human decides.
- Propose & approve — the agent prepares changes; a steward signs off before anything takes effect.
- Autonomous — for low-risk work, the agent acts inside policy boundaries, with a full audit trail.
Low-risk catalog hygiene can run autonomously while export-controlled data stays strictly advisory — in the same program.
Accountability is what makes autonomy safe
The goal isn't to limit what AI can do in governance; it's to make sure that whatever it does, you can always answer the auditor's question: who did this, and were they allowed to? Get that right, and AI becomes a force multiplier you can actually defend.