A pilot scope statement fixes, before an AI agent pilot starts, exactly what the agent will do, for which team, on what volume, at what depth, inside which limits, for how long, and what will stop it early. The pilot owner and risk owner sign it, and nothing outside it happens during the pilot.
Why write a scope statement for a pilot?
Pilots drift. A pilot that starts as "draft replies to billing emails for one team" ends as "send replies to most emails for three teams", one reasonable request at a time. By then nobody can say what was tested, what was approved, or whether the results mean anything. A pilot is the first small step on the path of train, shadow, certify, widen; scope creep skips the steps.
A scope statement prevents that by acting as a scope gate: anything outside it is either refused or waits for a written change. It also ties the pilot to the firm's limits. Each line should be traceable to the AI agent risk appetite statement, so the pilot cannot quietly test something the board has not accepted.
When should you use it?
Before the pilot starts, and again at any change to it. Pair it with the pilot success criteria checklist: the scope says what will be tried, the criteria say how you will judge it.
A pilot works best in one team with one process. Running a pilot of an AI agent in one team explains why.
How do you fill in each field?
- Objective is the question the pilot answers, not a hope. "Can the agent draft billing replies that agents send unchanged at least as often as templates?" is a question.
- In scope and out of scope list tasks and populations. Make the out-of-scope list longer.
- Depth for pilots is usually 1 or 2. A pilot at depth 3 needs evidence first.
- Volume caps the work, by day or in total.
- Data lists the classes the agent will see, which drives any data protection assessment.
- Stop conditions say what ends the pilot early, and who decides.
- Exit route names the decisions available at the end: stop, extend, move to shadowing at scale, or prepare for production. The last is covered in how to move an AI agent from pilot to production.
The template
| No. | Field | What to write |
|---|---|---|
| 1 | Pilot name and agent | Name, inventory ID |
| 2 | Objective | The question the pilot answers |
| 3 | Team and process | One team, one process |
| 4 | In scope | Tasks and customer or case types |
| 5 | Out of scope | Tasks and types excluded |
| 6 | Depth | Usually 1 or 2 |
| 7 | Limits | Amounts, volumes, hours |
| 8 | Data | Classes and systems |
| 9 | People | Pilot owner, reviewers, approvers |
| 10 | Duration | Start and end |
| 11 | Stop conditions | Events that end the pilot early |
| 12 | Appetite reference | Lines of the risk appetite relied on |
| 13 | Exit decisions | Options at the end and who decides |
| 14 | Sign-off | Pilot owner and risk owner |
A filled sample
A hypothetical energy retailer's billing team.
| No. | Entry |
|---|---|
| 2 | Can the agent draft replies to billing queries that handlers send with no more than minor edits, at least as often as with current templates? |
| 3 | Billing team B, email queue only |
| 4 | Bill explanations, payment date questions, meter reading queries |
| 5 | Complaints, debt and payment plans, customers on the priority services register, anything mentioning medical equipment, all other channels |
| 6 | Read billing account: 1. Draft reply: 2 |
| 7 | Up to 150 emails a day; business hours only |
| 8 | Customer name, account and billing history; no payment card data |
| 9 | Pilot owner: billing team B manager. Reviewers: six handlers. QA: billing quality lead |
| 10 | Six weeks |
| 11 | Any draft sent containing another customer's data; more than three complaints linked to agent drafts; data protection officer's request |
| 12 | Draft for a person, moderate appetite, ceiling 2 |
| 13 | Stop; extend by four weeks; move to shadowing across all billing teams. Decided by head of billing with operational risk |
| 14 | Billing team B manager; operational risk partner |
Row 5 excludes customers on the priority services register entirely. A pilot is the wrong place to learn how the agent treats people who most need careful handling.
Questions people ask
How long should a pilot run?
Long enough to see the full range of normal work, including month-end or other peaks. Four to eight weeks suits most processes.
Should a pilot ever let the agent act on its own?
Rarely. Pilots are for learning whether the agent's work is good. Acting on its own comes later, after certification on real cases.
Can we change scope during the pilot?
Only with a signed change to the statement. Record what changed and when, so results before and after can be read separately.