Scoping AI systems

Start from the decision people need

Write the decision or workflow the system must improve. Document intelligence, retrieval, copilots, and agents only earn a place when they change an outcome people rely on.

If you cannot say what “good” looks like for the people who use it every day, you are not ready to pick tooling.

Boundaries that belong in the brief

Lock data sources and retention, latency budgets, grounding rules when answers must stay factual, and what happens when the model is wrong or unavailable.

Name the systems you already run (identity, storage, ticketing, phone, CRM) and how the new path joins them. Integration surface is usually where calendars slip.

Someone has to run it

Agree who watches quality, cost, and incidents after launch. A pilot that nobody watches becomes a liability the week traffic arrives.

Accuracy claims and rankings matter less than clear answers on those points.

Where we have done this

KeytoZ work includes document and project intelligence on DataBonder, and AI phone answering (AnswerBug) for EVS7. See Work for the cases.