Only when the evidence says so, and only for a specific depth. An agent is ready when it clears the floor on every gate tested on your own past cases, its shadow results on live work match, and the operation around it is proven: named owner, staffed queues, a tested stop switch, and a fallback people can use.
Why is "ready" always ready for something?
Because readiness is per depth. An agent can be ready to draft customer replies for people to send long before it is ready to send them itself. Asking "is it ready?" without saying for what invites a vague answer. Asking "is it ready to draft replies at depth 2 for these query types?" invites a testable one.
So the first readiness decision is the agent's starting depth. Most agents should start lower than their sponsors hope, and rise on evidence.
What must be true before go-live?
| Test | Pass condition | Evidence |
|---|---|---|
| Performance on your own cases | Clears the floor on every gate, including refusal, escalation and staying in scope | Scored case results |
| Shadow on live work | Results match the test results; no new failure types | Shadow comparison report |
| Owner | A named process owner who accepts the role | Ownership record |
| Queues and approvers | Hand-over queues staffed; named approvers and cover | Rota and approval matrix |
| Limits | Enforced outside the agent; tested with cases above them | Test cases and verdicts |
| Stop switch | Used once in testing by the person who will use it live | Test record |
| Record | A single decision can be reconstructed from it | Reconstruction test |
| Fallback | Manual procedure current; staff know it | Fallback document and briefing |
| Approvals | Security, risk, data protection signed off | Approval records |
| Monitoring | Indicators and reporting set up | Monitoring plan |
The first row includes the cases where the right answer is "no" or "not me". An agent that scores well on routine cases but acts where it should refuse or hand over is not ready at any depth above drafting.
Who decides?
The process owner proposes go-live with the evidence; risk agrees; the accountable executive approves anything at depth 4. The owner should be in operations, because operations will live with the result; see why operations, not IT, should own agents in production.
What does this look like in practice?
Take a payments company deploying an agent to run KYB checks on new merchants: verifying the business, its owners and its activity. Testing on three hundred past applications shows the agent clears the floor on document checks and ownership tracing, and correctly holds every application involving a trust or a high-risk activity.
Shadow running for three weeks finds one new failure type: merchants registered in one country but trading mainly in another, which the test set did not include. The agent was treating them as low risk. Those cases are added to the test set, a hand-over rule is written, and the agent is re-tested.
The readiness review then checks the rest. The onboarding queue has named approvers and cover. The team lead has used the stop switch in testing. A sample decision is reconstructed from the record in a few minutes. The agent goes live at depth 2: it prepares each KYB file; an analyst decides. Widening criteria are written into the approval.
What happens after go-live?
The first week is part of readiness. Watch holds, edits, rejections and anything unexpected daily; running an agent's first week in production covers how. If the agent fails in production, its authority drops while the cause is found; see what happens when an AI agent fails in production. Moving from a contained pilot to wider use is a separate decision with its own evidence, covered in how to move an AI agent from pilot to production.
Where Pryme fits
In Pryme, Attest certifies each agent by replaying real cases from your operation against a floor across seven gates: grounded, refuses, depth, escalates, scope, adversarial and stable. Each must clear on its own. The certificate sets the depth, capped by its role's ceiling, and lapses after 90 days. More in deploying an AI agent: train, shadow, certify, widen.
Questions people ask
How long should shadow running last?
Long enough to see the full range of normal work, including periodic events like month-end. For many processes that is a few weeks.
Can an agent be ready if it fails some test cases?
Yes, if the failures are on cases it hands over correctly, or if the capability starts at drafting with people deciding. Not if it acts wrongly where it should refuse.
Who should sign the readiness review?
The process owner and risk, with security and data protection confirming their parts.