Start with one task
“Add AI” is too vague to build or test. Choose one user task: find the right procedure, draft a reply, or extract a set of fields. Write down what the person does today and what they must still approve after the feature exists.
Collect test examples
Collect synthetic or approved examples of ordinary requests, ambiguous inputs, missing information, and attempts to access an unauthorized record. For each, describe an acceptable response and a failure. Hold some examples back, so you can check a prompt change against cases it wasn’t tuned on.
Control access in your code
The AI shouldn’t decide what it’s allowed to see. Your code should limit searches and tools to what the current user can access. Documents can contain harmful instructions, so treat anything the AI reads as data. Require approval before changing business records or sending a message.
Measure usefulness and cost
Compare task completion, review effort, refusal behavior, latency, and cost per completed workflow against the current process. Record the model configuration and test date. A well-written answer can still be wrong, and a slow one can cancel out the time saved.
Decide what happens after the pilot
Before the pilot, write down the conditions for proceeding, revising, or stopping. A production scope needs ownership for monitoring, provider changes, data deletion, user feedback, and rollback. A prototype that works in testing can still make mistakes later, so keep checking it.
Further reading
Sources checked on 16 October 2026. This is planning advice, not results from a client project.