Knowledge
What controls should exist before AI is allowed to act?
Three, before anything else: rules stated in advance for what AI-started work may do, a person approving where those rules require it, and a record of what actually happened rather than only where things ended up. Without all three, you cannot answer the question that matters after an incident — what did it do, and who allowed it.
The three, in order
- 01
Rules, stated up front
What AI-started work may do, written down in advance rather than discovered from what it did.
- 02
Approval where required
A person reviews and approves before anything is applied, wherever your rules say so.
- 03
A record of every run
What was proposed, on what basis, who approved it, and what the system actually did.
- 04
Execution by the owner
The change is made by the system that owns it, so the record ends up where the information already lives.
Why a final state is not enough
Most systems record what a value became. That answers "what is true now" and nothing else. After an incident, the questions are different: what was proposed, what did the proposal rest on, was it approved, and by whom.
A record of who or what acted, on which record, doing what, and how it looked before and after answers those. A changed field does not.
The failure this prevents
The expensive failure is not an AI system doing something dramatic. It is an AI system doing something small and plausible, repeatedly, that nobody notices until the aggregate shows up somewhere else — and then finding there is no record of why any of it happened.
Questions buyers ask
- Does approval on everything make AI useless?
- It would, which is why the policy declares where approval is required rather than requiring it everywhere. The judgement is which changes are reversible and which are not.
- Is this the same as an audit log?
- An audit log is part of it. The other parts are the policy that says what was permitted and the approval record that says a person agreed.