Start with the bottleneck, not the model
The strongest starting point is a recurring task where people spend substantial time searching, sorting or drafting. Examples include qualifying an enquiry, preparing a reply or bringing information together.
Ambiguous, high-risk decisions make poor first use cases because apparent automation quickly becomes extra review work.
- The task occurs regularly in a similar form
- Suitable source data is available and current
- A wrong suggestion can be recognised and corrected
- Value can be measured through time, quality or throughput
Put sources and boundaries in the interface
Users should see what an answer is based on and what status it has. A draft is not a sent message, a recommendation is not an approval, and a summary does not replace its source.
These boundaries belong in the product through source links, editing, roles and precise action labels—not only in an internal policy.
Design the human handoff
A dependable assistant recognises when it should stop. Missing data, sensitive content, unusual cases or low confidence can trigger a handoff.
The handoff must preserve the conversation, relevant sources and the reason for escalation. Otherwise the person starts again and the promised efficiency disappears.
Clarify privacy and operations early
Before prototyping, establish which data is processed, where it is stored and who may access outputs. Personal or confidential content requires a deliberately chosen architecture.
Ongoing quality matters too: which examples are reviewed, how are problematic outputs reported, and who approves new sources or rules?
A useful first pilot
A four-to-six-week pilot should cover one narrow workflow with real cases. Measure handling time, correction effort and the quality of handoffs.
Expand only after that core works reliably. This protects the budget and builds trust with the people who use the system every day.