Check whether AI is necessary
Start with the decision or repetitive task you want to improve. Stable rules and structured inputs may be better served by ordinary automation. AI is worth evaluating when the inputs vary, such as document wording, but the outcome can still be checked.
Map sources and permissions
List where the source material lives, who owns it and who may read it. Record how updates and deletions reach the application. A document assistant needs permission checks at retrieval time, not just a login screen. Use synthetic or approved samples during discovery.
Build an evaluation set
Collect representative questions or tasks with expected answers, source passages and examples where the system should decline. Include outdated documents, missing evidence and conflicting instructions. Agree how errors will be reviewed before making the workflow automatic.
Keep actions reviewable
Separate drafting an answer from changing a record or sending a message. Define which actions need approval, how failed actions are retried and how to stop the workflow. Compare operating cost, latency and correction effort against the current process.
Bring this to your project brief.
- Approved sample inputs and expected outputs
- Source owners and access boundaries
- An evaluation set and human-review rules
Use general descriptions or anonymized examples. Do not put credentials, personal records or confidential documents into the brief.
Prepare your project brief