Name the owner of the operational result
The sponsor who approves the budget and the person who accepts daily output may be different people. Identify both. Write down who can change the process and who decides when an exception needs a person. A delivery team cannot resolve conflicting business instructions through model selection.
Choose the workflow boundary
Describe the trigger and the point at which work is complete. List the applications and handoffs in between. Mark steps with formal rules separately from steps requiring interpretation. This creates a useful unit for estimation and avoids an open-ended request to “add AI everywhere”.
Record the data and access path
For every source, identify an owner, interface and permission model. Note whether examples can be anonymized and whether a test environment exists. State the proposed processing locations and external services for review. An EU supplier, an EU database and an EU-only processing chain are different things.
Define acceptance before implementation
Collect representative examples and difficult exceptions. Decide what a reviewer must see to accept a result and what makes an action unacceptable. Include unavailable systems and incomplete information. Keep evaluation examples separate from the examples used to tune the implementation.
Specify ownership after the release
Someone must update documents, manage API changes, investigate failures and approve changes to agent behavior. Ask for the operating instructions and test evidence needed by that team. Clarify source-code access and deliverable rights in the agreement rather than assuming a generic ownership policy.
Agree how the engagement will run
For a European or US buying team, record the contracting entity, delivery language, review participants and meeting overlap. Ask which evidence is available for your vendor review. Separate requested certifications from verified credentials, and distinguish project-specific commitments from capabilities still being evaluated.
Bring these to the first meeting
- A one-paragraph business objective
- The applications and data owners involved
- Two typical examples and one difficult exception
- Known access, procurement and timing constraints