Three things must be true first: a named operations owner has accepted the agent, it has been certified on real cases from your operation rather than the pilot's hand-picked ones, and it starts in production at a lower depth than it showed in the pilot. Production is a change in who answers for the agent, not just where it runs.
Why is the step from pilot to production so hard?
A pilot is built to show that something can work. It usually runs on clean cases, with the people who built it watching every output and fixing problems quietly. Production removes all three supports at once: the cases get messy, the builders move on, and nobody is watching each output.
Most pilots that stall do so here, because the evidence gathered in the pilot answers the wrong question. "Can the agent do this?" is settled. "Can we let it do this, every day, under an owner who did not build it?" is not.
What has to be in place before the move?
- An owner in operations. The person who runs the process owns the agent, with the authority to narrow or stop it. Our view on why operations, not IT, should own agents in production sets out the reasoning.
- Success criteria written before the pilot ended. If they were written afterwards, they were written to fit. The pilot success criteria checklist is a useful test of whether yours were specific enough.
- A scope that matches the pilot's. If the pilot covered one product line in one team, production starts there too. The pilot scope statement template gives you a fixed reference to hold the line against.
- Certification on your own cases, scored against floors set by the owner, not by the vendor or the build team.
- A rollback route, tested, so the work can go back to people in minutes.
What does the agent do in production, and what stays with a person?
| In the first weeks of production | The agent | A person |
|---|---|---|
| Reading cases and gathering facts | Yes, depth 1 | Spot-checks a sample |
| Drafting replies, notes or entries | Yes, depth 2 | Approves and sends each one |
| Routine actions within limits | Only if certified for that capability | Reviews the held queue daily |
| Anything that moves money or is hard to undo | No | Decides and acts |
| Changing the agent's scope or depth | No | Owner, on evidence |
Starting a notch below the pilot feels like a step backwards. It is the opposite. In the pilot, people were catching problems you never saw. In production, the held queue and the draft edit rate show you what those people were catching.
What does a good move look like?
Take a regional bank's payments operations team that piloted an agent to investigate failed outbound transfers. In the pilot it gathered the payment record, checked the beneficiary details and drafted a repair instruction, and the builders sent most of them directly.
For production, the payments operations manager becomes owner. The agent is certified on 300 failed transfers from the last quarter, including returned payments and name-mismatch cases the pilot never saw. It clears every gate for reading and drafting, and the owner sets it to depth 2: it drafts the repair, a person sends it. After six weeks of filed verdicts showing a low edit rate and no scope breaches, the owner asks for a fresh certification at depth 3 for one repair type only.
Which controls make the transition safe?
- A dated handover. The build team signs the agent over, and the owner signs to accept it. Both are filed.
- Production certification, separate from any pilot results, with its own expiry.
- A held queue with a named reviewer and a daily review slot, not "whoever is free".
- Drift monitoring from day one, compared against the certification baseline.
- One change at a time. Do not widen scope and depth in the same week, or you cannot tell which change caused a problem.
The full sequence of train, shadow, certify and widen is in our guide to deploying an AI agent. If the pilot itself was not set up to produce useful evidence, start with how to run a pilot of an AI agent in one team.
Where Pryme fits
In Pryme, depth is earned, not configured. The agent's role sets a ceiling and Attest evidence sets the depth, and the lower wins. A pilot result does not carry into production: the agent is certified on the owner's cases, the certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day.
Questions people ask
How long should the first production phase last?
Long enough to see a full business cycle, such as a month-end or a billing run. For most processes that means four to eight weeks before any request to widen depth.
Can the pilot team keep running the agent?
They can support it, but the owner must sit in operations. A builder who also approves the agent's work is marking their own homework.
What if the pilot showed the agent ready for depth 3?
Pilot evidence was gathered under supervision. Start production at depth 2 and let filed production verdicts support the move, through a fresh certification.