PrymeIntelligence

Regulated autonomy·Playbook

How to decide which actions an AI agent may take without a person

Decide which actions an AI agent may take without a person by running each through six tests: reversal, effect, value, sources, evidence and volume.

An action may run without a person only if it passes six tests at once: your team can undo it alone, it has no legal or similarly significant effect on someone, it stays within limits, it relies on trusted sources, the agent has passed it on your own cases, and an error would not spread widely before anyone noticed.

What is this decision, and who owns it?

Every capability an agent holds sits somewhere between "a person does everything" and "the agent acts alone". This playbook is about the last step: the actions the agent may complete with no human in the loop. Our guide to autonomy in regulated industries argues that this is a question about reversal first. The other tests narrow it further.

The process owner proposes. Risk and compliance challenge. The accountable executive, the person who would answer to a regulator for the outcome, signs. That signature should name the capability, the limits and the expiry date.

Which six tests must an action pass?

  1. Reversal. Can your team undo it alone, completely, with a tested procedure? If not, stop here. The list of which agent actions are irreversible is a good starting check.
  2. Effect on a person. Does it produce a legal or similarly significant effect on someone, such as a refusal of credit, a change to their contract, or a restriction on their account? Under GDPR Article 22, decisions of that kind based solely on automated processing carry specific conditions. Treat any such action as needing a person unless counsel has confirmed otherwise.
  3. Value. Is it inside the customer's own limits, per action and per day?
  4. Sources. Does every fact it relies on come from a signed or sealed source, not a declaration or an outside claim?
  5. Evidence. Has the agent been certified on this exact capability against your own past cases, and is the certificate in date?
  6. Volume. If the agent got it wrong on every case for a day, how many customers would be affected, and would you notice?

How do you make the decision, step by step?

  1. Person, process owner: starts from the excluded list. Some actions no agent should take at all; decide those first, using how to decide which actions no AI agent may take. Evidence: an exclusions list.
  2. Person, process owner: runs the six tests on every remaining action. Record a yes or no with a reason for each. Evidence: a test sheet per action.
  3. Person, data owner: confirms the data each candidate needs. An action that can only run by seeing data the agent should not see fails here. The method is in how to decide which data an AI agent may see. Evidence: data approval per capability.
  4. Person, operations: attaches a tested undo procedure. The undo procedure template covers the fields. Evidence: procedure with a test date.
  5. Agent, depth 1 and 2: runs in shadow, then drafts. It proposes the action on live cases; a person carries it out. Evidence: drafts and how often people changed them.
  6. Platform: certifies the capability on past cases. Evidence: a scored run against the floor.
  7. Person, accountable executive: signs the autonomy decision. Capability, limits, sampling rate, expiry. Evidence: the signed decision.
  8. Agent, depth 3: acts within bounds. Evidence: allowed verdicts, each citing the bounds in force.
  9. Person, quality reviewer: samples completed actions. Evidence: sample results fed back to the owner.
Step The agent does Depth Stays with a person
Exclusions Nothing none Owner lists actions no agent takes
Six tests Nothing none Owner records each test and reason
Data check Nothing none Data owner approves
Undo procedure Nothing none Operations writes and tests
Shadow and draft Proposes, then drafts 1, then 2 People act on drafts
Certify Replays past cases 1 Owner reviews failures
Sign Nothing none Accountable executive signs
Act Completes the action within bounds 3 Sampling and exceptions

Which controls keep the decision honest, and what evidence do they leave?

Every test recorded. A decision with a reason per test can be challenged and defended. Evidence: test sheets.

Lowest result wins. An action that passes five tests and fails one does not run alone. Evidence: the signed decision lists every test.

Expiry. Autonomy granted on old evidence describes a past agent. Tie the decision to the certificate's 90-day life. Evidence: expiry date on the decision.

Automatic step-down. A failed certification run or a breach drops the capability back to read the same day. Evidence: step-down events.

Sampling after go-live. Evidence: sample size, findings, and actions taken.

What should you measure?

  • Sample error rate on completed actions, counted per hundred sampled.
  • Undo procedures used, and whether each worked.
  • Complaints linked to unattended actions.
  • Time from case open to close before and after, to check the autonomy is worth its oversight cost.
  • Tests that changed answer at renewal. A test that flips from yes to no means something moved.

Where does it usually go wrong?

Reversal is the only test. A reversible action applied to every customer in a day is still an incident. Volume and effect matter. The reasons are set out in whether a reversible action is always safe for an agent to take.

The tests are run once. Systems, sources and policies change.

Draft is skipped. Moving straight from shadow to acting removes the period where people see real drafts and correct them.

Sign-off sits too low. A team lead approves autonomy that the executive would not have. Match the signature to the accountability.

Questions people ask

Can an action pass for one customer segment and not another?

Yes. The effect test often differs: a fee waiver for a business account is not the same as one for a customer in financial difficulty. Split the capability by segment and decide each separately.

How often should the decision be revisited?

At every certificate renewal, and after any incident, policy change or new data source affecting the capability.

What if the business wants autonomy before the evidence exists?

Run the agent at draft and collect the evidence. Drafting still saves most of the effort, and it gives you the cases you need to certify.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.