PlaybookMoving an AI pilot into governed production
For AI program leads, platform owners, and the product and risk owners who sign off a go-live
21 November 2026
An approved AI use case is not yet a production service. Between the two sits a body of work that most programs underestimate: the controls that make the service defensible, the gate that decides it is ready, and the hand-over that keeps it running once the project team moves on. This playbook sets out that work in the order it needs to happen.
1. What good looks like
A use case in governed production has a named owner, its policy enforced on every request, its cost attributed to the business that benefits, a record of what it did kept for as long as your obligations require, and a scorecard someone with authority reviews on a schedule. That is the baseline the service improves from, and the evidence on which to expand, hold or retire it.
2. Readiness before you start
Most delays are decided before the work begins. Check four things and do not start until they are true. Scope: a one-paragraph description of the use case with what is in and out of bounds written down. Data: access to every system the use case touches, tested rather than promised. Ownership: three names, one each for product, technology and risk, with time in their calendars. Baseline: the business measure the service is meant to move, measured before anything changes. A vague scope, untested access or an ambiguous owner will each send the work backwards within a fortnight.
3. Reuse what is common, configure what is yours
Speed comes from not designing what has been designed before. Policy templates by risk tier, trace and evaluation schemas, the structure of the cost ledger, the evidence format agreed with audit, the scorecard template and the common runbooks should all be reused from a standard pattern. What genuinely has to be configured for your organisation is narrower: your data sources and access boundaries, your risk appetite and approval paths, your budget caps and cost allocation, your owners at each gate, and the quality thresholds specific to the use case. If a team finds itself writing the evidence format from a blank page, the timeline has already slipped.
4. The four controls
Stand up four controls around the use case, all of them live before any real user touches it. Policy enforcement: the rules applied at a single control point on every request, with an exception path that has a named approver. Tracing and evaluation: every interaction traced end to end and scored against the quality definition for the use case. Cost attribution: every unit of compute assigned to the use case and business unit, with a cap enforced when the budget is reached. The evidence record: the request, the policy decision, the response and any human review, retained under a policy agreed with audit and legal.
5. Integrate, do not replace
The controls should sit on top of the platforms you already run, not replace them. Most modern AI platforms already emit the signals you need; the work is connecting them to your approvals, your monitoring and your finance structure. Teams that use a go-live as the moment to swap out the underlying stack lose weeks and end up where they started.
6. The go-live gate
Go-live is a gate with pass criteria, not a date. Policy enforcement is verified against attack tests on all traffic. Every interaction is traced and scored. All cost is attributed and the cap is enforced. The evidence record holds what it should. The scorecard is published with its baseline. The runbook is approved and an on-call rotation exists. Run the gate as a meeting, not an email: product, technology, risk, audit, finance and the delivery lead in one room for an hour. If anyone says not yet, it does not ship that week.
7. Hand over to operations
The project ends and the service begins, and the hand-over is where many services quietly decline. Archive a go-live pack: the policy configuration, trace dashboards, cost map, evidence format and runbook. Book the monthly scorecard review with the executive sponsor and the quarterly review ninety days out. Test the incident runbook with a live drill, not a read-through. Formally transfer the on-call rotation to the team that will run the service. If the go-live pack cannot be produced cleanly, the work is not done, even if users are already on the system.
8. What usually goes wrong
Three things reliably stretch the timeline: unclear data ownership, so nobody can approve access to the source the use case needs; no named business owner to sign the go-live; and scope creep, most often a second use case added part-way through. The first two are readiness failures and should be caught before the work starts. The third is a line that has to be held. When the work runs long, it is almost always one of these rather than the technology.
The test of the whole playbook is the first monthly review after go-live. If the owner can show the scorecard, the cost trend and the evidence record without anyone assembling them by hand, the service is in governed production. If the meeting needs a week of preparation, it is still a pilot.