Technical due diligence questions for an AI implementation partner

Trace a request across the system

Start with the user’s request and follow it through identity, retrieval, model processing, tools, and the destination application. Identify where business rules are enforced and where state is stored. Ask what prevents a request from reaching the wrong account or record. A system diagram is useful only when the delivery team can connect its boxes to observable behavior and explain who owns each boundary.

Inspect the interface around the model

For an OpenAI implementation, ask which runtime is proposed and which application responsibilities remain outside it. OpenAI documents distinct agent runtime approaches; the vendor should justify its choice against your needs. Request examples of tool inputs, validation, and error results. The design should explain how an uncertain interpretation becomes a reviewable proposal before it can trigger an important business action.

Reference: OpenAI: Agent runtime options

Ask for failure evidence

Choose a few difficult cases such as an unavailable source, stale record, denied permission, and interrupted action. Ask how the team would test them and preserve the results. A vendor need not reveal another customer’s private implementation to demonstrate engineering discipline. It can explain its testing method, show appropriately sanitized artifacts, or perform an agreed exercise using material your organization has approved.

Assess the next maintainer’s position

Review how a new engineer would find configuration, run checks, investigate a failed case, and understand release changes. Ask which dependencies create continuing vendor reliance. Agentix offers custom agent development, but fit should be established through a specific brief and verifiable artifacts. Capture unresolved questions in the discovery scope so they do not quietly become assumptions that survive into deployment.

Reference: Agentix (publisher): Agentix services

Common questions