THE DIRECT ANSWER
An AI system handover is complete when the internal owner can run the workflow, check its outputs, handle failures and manage access without relying on the builder. Require documented assets, evaluation results, cost and support responsibilities, and a witnessed acceptance test before signing off the agreed scope.
01
Agree the acceptance test before development starts
Define the business task and its boundaries in the scope. Identify the internal operator, the person who approves material actions and the person who maintains the system. They can be different people. Set acceptance criteria for routine work, difficult cases and recovery.
Write what the system must do when it cannot produce an acceptable result. Pausing for a person, keeping the original input and showing the reason can be the correct outcome. A polished demonstration that only includes straightforward examples does not establish readiness.
02
Require these six parts of the handover pack
Keep sensitive access material in the company's approved secure system. The general operating guide should reference that location without exposing credentials. Record third-party limits so the team understands what it owns, what it licences and what depends on an outside service.
| Part | What to receive | How to verify it |
|---|---|---|
| Access and assets | Account owners, repositories or exports, credentials transfer process and licence inventory. | The company owner can access the required systems using company-controlled accounts. |
| Operating guide | Inputs, steps, outputs, approvals and named exception owners. | The internal operator completes a normal case from the guide. |
| Evaluation record | Test cases, acceptance criteria, observed failures and known limits. | The owner repeats an agreed sample and checks results against the criteria. |
| Recovery guide | Stop procedure, manual fallback, incident routing and restore steps. | The team handles a controlled failure without builder intervention. |
| Cost and maintenance | Software dependencies, usage assumptions, billing owner and update responsibilities. | The owner can explain ongoing costs and who responds when tools change. |
| Training and sign-off | Working sessions, operator practice and unresolved-item register. | The accountable owner records acceptance or the specific remaining work. |
03
Run a witnessed acceptance session
The builder should observe instead of quietly fixing the test. If the owner needs help, log the missing step and repeat the relevant test after training or documentation improves. Agree any controlled exceptions explicitly before signing off.
- Ask the internal owner to start the workflow with an ordinary approved input.
- Ask them to check the output against the source records and acceptance criteria.
- Introduce a missing input and confirm that the workflow uses the agreed exception path.
- Simulate an unavailable connection and ask the owner to use the recovery guide.
- Confirm that approval controls prevent unauthorised external actions.
- Have the owner locate the usage report, support contact and change log.
- Record the result, observed limits, unresolved work and the final acceptance decision.
04
Assign responsibility after launch
AI workflows can change when models, prompts, source data, tools or operating rules change. Maintain a version record and repeat relevant tests before approving a change. Monitor failure and review rates along with usage. Agree a person who can pause the workflow when quality drops.
The NIST AI RMF Playbook offers voluntary actions for managing AI risk through design, deployment and use. Its continuing-risk approach supports an operating review after handover. The checklist here is Tenth's practical planning aid, not a NIST certification.
- Name the maintenance owner and the conditions that trigger a review.
- Choose routine sample checks and a process for investigating failed outputs.
- Record who approves system changes and when evaluation must be repeated.
- Agree support hours, response expectations and commercial scope if outside support continues.
05
Illustrative acceptance test: a weekly reporting assistant
A reporting assistant collects approved business data and prepares a weekly review draft. Acceptance should require correct source references, visible missing-data labels and human approval before distribution. Test a normal week, a missing data source and a conflicting metric definition.
The internal owner must produce the draft, find the evidence for a number, correct an exception and use the manual fallback. If the builder remains the only person who can repair a broken connection, the system still needs a defined maintenance arrangement or further handover work.
Illustrative example. The agreed tests depend on the actual system and business requirements.
PUT THE GUIDE TO WORK
Define what your team must be able to run.
Tenth builds AI systems in approved scopes with a named internal owner, documentation, training and a completion test. Stewardship after implementation is optional.
Discuss an AI buildAbout this guide: Tenth provides growth management and AI engineering for private equity firms and their portfolio companies. The checklists are editorial recommendations. Adapt them to the company's evidence and agreed responsibilities. Send a correction or discuss the work.