Evidence Before Autonomy
A practical ladder for granting AI agents more authority without losing the audit trail — and the evidence each rung should require.
The question leaders ask most about agents is how much to trust them. It is the wrong question. Trust is a feeling; what a board, a regulator or an auditor will ask for is evidence. The better question is: what would we need to be able to show before we let an agent do more?
A useful way to answer it is a ladder. Each rung grants the agent more authority, and each rung has an entry requirement written in evidence rather than confidence.
The five rungs
- Observe. The agent reads the work and proposes what it would do. Nothing changes. Evidence to move up: its proposals are compared against what people actually did, and the gap is measured.
- Assist. The agent prepares the work; a person executes it. Evidence to move up: a record of every draft, every human edit and why.
- Act with approval. The agent executes after a named person approves each action. Evidence to move up: approval logs that show who approved what, and how often approvals were overridden.
- Act and report. The agent executes within a defined scope and reports afterwards; people sample and review. Evidence to move up: sampling results, exception rates and a clean trail from request to outcome.
- Act within bounds. The agent runs the process inside hard limits — scope, time, systems, value — with automatic escalation outside them. Evidence to stay here: continuous logging, periodic review, and a named owner accountable for the agent.
Trust is a feeling. What the auditor will ask for is evidence.
Three rules that keep the ladder honest
Every agent has an owner. A person, not a team alias, who is accountable for what it does and who can switch it off.
Scope is explicit. What the agent may touch, for how long, and up to what value is written down before it runs — not inferred from what it happened to be given access to.
Moving down is normal. An incident, a policy change or a new system should drop an agent a rung until the evidence is rebuilt. Programs that treat demotion as failure stop reporting problems.
Where to start
Take the process you would insource first and write down which rung it is on today and what evidence would move it one rung up. If nobody can say where that evidence would be stored, that is the first project.
Part of Issue 01, “The Insourcing Era.” Governance leaders: this is a standing topic at InMag Roundtables. Request a seat.
