Three things need training: when to pause an agent into draft-only mode, who may do it, and how the team absorbs the surge of drafts that follows. Agree the triggers and named decision-makers in advance, rehearse with a timed drill, and practise the return to normal, which needs evidence rather than relief that the incident is over.
What is draft-only mode, and who owns the decision?
Draft-only mode restricts an agent, or one of its capabilities, so that it drafts and a person sends. Work it would normally execute at depth 3 or 4 arrives as drafts or held items instead. It is a narrowing the business applies through its own limits, and it can be switched on quickly. It differs from what happens when a certificate fails: a run that clears no depth drops the agent to depth 1 the same day, automatically. Draft-only is a choice made by people.
The decision belongs to named roles, not to whoever is online. A typical list: the head of the function, the on-call operations lead, and the risk or incident manager. Any one of them may switch to draft-only; returning to normal needs the function head and the builder to agree. The incident manager owns the training.
When should a team pause an agent?
Teach the triggers as a short list with examples:
- Suspected manipulation at scale. Several planted instructions found in incoming documents within hours.
- A source turns out to be wrong. The policy library holds an outdated fee schedule, or a data feed is stale.
- A policy change not yet reflected. The business changes its refund policy at short notice and the agent's brief has not caught up.
- A spike in complaints about the agent's output. Several customers reporting the same wrong statement.
- Doubt about an upstream system. The ledger or settlement rail reports an issue.
The test to teach is simple: if executed actions could cause harm that a person would catch, pause. Pausing is cheap and reversible; a wave of wrong executed actions may not be.
The steps
- Person (incident manager): publishes the triggers and the named decision-makers, with contact routes out of hours.
- Person (team): learns the triggers and how to raise one, in a short session.
- Person (any reviewer): raises a trigger to a named decision-maker with the evidence seen.
- Person (decision-maker): switches the capability to draft-only and records why.
- Agent, depth 2: drafts everything it would have executed for the paused capability; held items continue as before.
- Person (team lead): restaffs: moves reviewers onto the affected queue and tells customers about any delay.
- Person (builder and function head): fix the cause, and recertify the capability where needed.
- Person (function head and builder): agree the return to normal on evidence, such as a clean run on recent cases, and record the decision.
- Agent, depth 1: logs the switch, every draft decision and the return.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Trigger raised | Shows evidence from the record | 1 | Reviewer raises; decision-maker decides |
| Paused capability | Drafts instead of executing | 2 | Reviewers send, edit or decline |
| Held items | Continue to hold | held | Named reviewers decide |
| Withheld actions | Remain withheld | withheld | Unchanged by the pause |
| Return to normal | Resumes the certified depth once restored | 3 | Function head and builder agree |
How should the drill run?
Run it unannounced in working hours. The trainer seeds evidence of a trigger, for example three supplier invoices in an hour carrying the same planted instruction. Start a timer.
Score four times: from first sight to a trigger being raised; from raising to a decision-maker switching to draft-only; from the switch to reviewers being moved onto the queue; and from the cause being fixed to a recorded return. Then debrief. Most teams find the slowest step is reaching a decision-maker, which is why the named list and contact routes matter more than any slide.
Which controls matter, and what evidence do they leave?
- Named decision-makers. Evidence: the published list and the name on each switch.
- Recorded reason for the pause. Evidence: the reason against the switch in the record.
- Evidence-based return. Evidence: the test results and both signatures on the return.
- Drill results. Evidence: the four times and the debrief notes.
- Append-only record. Evidence: everything the agent drafted and every human decision during the pause.
What should you measure?
Time from trigger to switch in drills and real incidents. Time to restaff. Draft backlog during the pause and how fast it cleared. Returns made without recorded evidence, which should be none. And incidents where a pause would have helped but none was called, found in later review.
Where does it usually go wrong?
Nobody feels entitled to pause. Reviewers see the problem and wait for someone senior. Make raising a trigger an expected duty, and thank people who raise false alarms.
The team pauses but cannot cope with the drafts. Plan the restaffing before the incident, including who stops other work.
The agent returns to full depth because the incident "feels over". Return on evidence.
Different sectors face the question of readiness in different ways; see whether hospitals can use agents before staff are trained and whether telecoms operators can. A short definition is in what draft-only mode is, and fuller drill designs in running an incident drill for the finance team. The guide to training teams for AI agents covers the wider programme.
Where the Academy fits
Pryme Intelligence Academy's Operator track includes a module on when something goes wrong, and the practicum asks for real review decisions with no missed red line. Module tests are set in the learner's own industry and role. Certified Pryme Operator is renewed each year through an update module and a short quiz, which keeps incident knowledge current as the product changes.
Questions people ask
Should customers be told when an agent is in draft-only mode?
Tell them about any delay that affects them, in plain terms. They do not need the internal detail, but they should not be left waiting without explanation.
Can draft-only mode apply to one capability only?
Yes, and that is usually better. Pausing only the affected capability keeps the rest of the work flowing.
Does draft-only mode affect the agent's certificate?
The pause itself is a business decision and does not change certification. If the cause was a fault in the agent's behaviour, the capability should be recertified before it returns to its previous depth.