When a production AI engagement ends, there is exactly one artifact that determines whether the system survives the consultant's exit: the handover document. Most engagements do not produce one. The system runs for nine months and then quietly degrades, because the knowledge of how it was built lives in an inbox the consultant no longer reads.
The handover document that makes a deployment durable has seven sections.
- Architecture diagram. Source systems on the left, model in the middle, destination systems on the right. Every integration named with its connection method, authentication scheme, and error handling pattern.
- The translation layer. The spreadsheet, document, or decision matrix that converts business rules into model-grade specifications. The artifact every future change is graded against.
- Confidence framework. Threshold per workflow, defended by the test data, with the exception path, the SLA, and the review cadence written down.
- Operational runbook. What to do when the model degrades, when an integration breaks, when an exception queue grows beyond SLA, when the audit team asks for evidence. Written for the operator who will inherit it, not the engineer who built it.
- Audit trail map. Where every decision is logged, in what schema, retained for how long, accessible by what query. The auditor's first question every year.
- Disaster recovery guide. What happens if the cloud environment dies, the vendor changes their API, or the model returns nonsense for forty-eight hours. Specific, named, with the escalation path included.
- Source code asset packaging. All custom code, prompts, integration scripts, and configuration files transferred to a repository the client owns, with the commit history intact.
In a two-year engagement involving a practice management platform integration into Sage Intacct, AWS RDS environment build-out, custom API transformation pipelines, and an extensive sandbox testing regime, the final deliverable was a handover document running over fifty pages across these seven sections, packaged for the client's incoming CTO. The CTO inherited a system, not a consulting dependency.
The test for whether your consultant is producing this document is simple. Ask: if you disappeared tomorrow, could my team operate and modify the system in six months? If the answer is anything other than yes, you are not buying production AI. You are renting access to one person's memory, and the meter is running.
A real handover document costs the consultant time and protects the client's investment. A missing handover document costs nobody anything until the consultant moves on — and then it costs everything.
