Subcontracting an AI integration: brief, acceptance and handover

An agency can have a project under way and still need a specialist for one AI integration. The assignment becomes difficult when responsibility for “the AI part” has no clear boundary. Identify a module that can be delivered and reviewed separately before bringing another engineer into the project.

The agency should retain project direction while the specialist has access to the decisions needed for implementation. Assign a technical contact rather than relying on forwarded messages to carry the whole brief.

Define the module and its dependencies

Describe what the module receives, what it returns and who uses the result. A screen can explain the interface, but it does not specify the data contract.

List what the agency will provide: the repository, a test environment, service documentation and examples suitable for sharing. If those inputs do not exist, agree who will prepare them. Dependencies belong in the scope as well as the code to be written.

When investigation is necessary, agree its price and deliverable. A screenshot alone may not provide enough information to estimate an unfamiliar integration.

Decide how the client is involved

Agree who can speak to the end client, which decisions need agency approval and how each participant is introduced. These arrangements should be clear before the first client meeting.

Technical review and business acceptance may have different owners. The agency might approve code while the client checks whether the result fits its process. Decide how disagreements will be resolved and who can authorize a change in scope.

Write acceptance examples

Consider a module that drafts a response from client documents. The table below is a hypothetical scope example.

Adapt these cases to the product. Agree latency and cost limits against known test conditions. Results from an unrelated demonstration cannot establish those limits for your integration.

Handover through the working repository

Ask for reviewable changes with documentation close to the code. Include how to run the integration, configure permissions and repeat the acceptance tests. Credentials should use the agreed secure mechanism and stay out of handover documents.

An agency engineer should be able to execute the checks and locate known limitations. Questions raised during that attempt are part of reviewing the handover.

Close the delivery explicitly

Record the accepted version and any remaining limitations. Separate defects in the agreed behavior from requests for new functionality. Support, maintenance and later provider changes need explicit terms.

Pharosyne can work with agencies on LLM integration, RAG and software architecture. Describe the module you need covered, what is already decided and who will review delivery. That gives us a basis for scoping work your team can take over.