Read what its job needs, and draft only where it has already been certified to draft, with a person sending every draft. Nothing that moves money, changes a customer's terms or is hard to undo, however good its test results. Day one is when you learn how it behaves on live work, and its authority should reflect that.
Why start so low if the tests went well?
Because tests and live work differ in ways you cannot see in advance. Live cases arrive with fresh data quality problems, new customer wording and timing you did not replicate. Reviewers are also new to the job of checking an agent. On day one, both the agent and the people around it are being tested.
Starting low costs little. An agent at depth 1 or 2 still does useful work: it gathers facts, prepares cases and writes drafts that save time. What it does not do is take an action that has to be unwound if it was wrong.
What does day-one authority look like?
| Kind of action | Day-one verdict | Why |
|---|---|---|
| Reading permitted systems and documents | Allowed | Needed for any work; nothing changes |
| Producing summaries, checklists and reports for staff | Allowed | Internal and easy to check |
| Drafting customer replies, notes or entries | Allowed only if certified at depth 2; a person sends | The person is the control |
| Routine updates within bounds, such as tagging or routing | Held, unless certified at depth 3 and the owner agrees | Wait for production evidence |
| Payments, refunds, limit changes, account closures | Refused at this depth | Hard to undo; earn it later |
| Actions withheld at the platform floor | Withheld, as for every agent | Not a day-one question at all |
"Allowed" here has a precise meaning: the action passed every bound in the chain, not just that someone approved the agent in principle. Our glossary entry on what allowed means sets out the distinction.
What does day one look like in practice?
Say a commercial bank's operations team switches on an agent to support daily cash reconciliations. It was certified to read the ledger and bank statements and to draft a list of unmatched items with suggested causes.
On day one it reads both sources each morning, produces the unmatched list, and drafts a short explanation for each break. An analyst checks every item and posts any adjustments. At the end of the day the agent runs its read-only checks and reports the day's matching rate, the kind of task described in running end-of-day operations checks with an AI agent.
It does not post adjustments, even small ones, though certification showed it could suggest them accurately. After four weeks of filed verdicts, a low edit rate and a clean month-end, the owner asks for fresh certification at depth 3, with posting limited to one type of adjustment: reversing timing differences under a set amount.
Which controls matter most on day one?
- Owner: confirms the agent's depth before go-live, in writing. Evidence: a filed depth setting.
- Reviewer: works the held queue at set times each day, using a routine like working through the agent's held queue each day. Evidence: time from hold to decision.
- Team lead: samples the agent's allowed work, not only its drafts. Evidence: a scored sample.
- On-call owner: is reachable for the whole working day, with authority to drop depth. Evidence: the rota.
- Owner: holds a short review at the end of day one and day five. Evidence: notes with any changes.
What should it never be allowed to do?
Some actions are not a question of when. They stay out of reach at every depth because the consequence is too large or the evidence can never be strong enough. The list is in what an AI agent should never be allowed to do. On day one, the useful discipline is to check that the agent's bounds already reflect that list, rather than relying on its brief.
Day one for the first agent also sets the pattern for later ones; see scaling from one AI agent to many. The full sequence is in our guide to deploying an AI agent.
Where Pryme fits
In Pryme, depth is earned, not configured. The role sets a ceiling, Attest sets the depth from evidence on your cases, and the lower wins. Every action meets the five bounds at once, and the verdict names which bound decided, so on day one you can see exactly why each action was allowed, held or refused.
Questions people ask
Can an agent start at depth 3 if the process is low risk?
It can, if certification supports it and the owner agrees. Even then, many owners start a step lower for a week to see live behaviour before relying on it.
Should the agent tell customers it is new?
Customers should be told when they are dealing with an AI system where rules or your policy require it. Its newness is an internal matter; the safeguards are what protect the customer.
How long should day-one limits last?
Until production evidence covers a full business cycle. For most processes that is four to eight weeks, after which a fresh certification can support a wider depth.