all work
ai systemsB2B software2025
A support assistant that drafts replies from the docs it can cite.
A support team was drowning in repeat questions that the documentation already answered. We built an assistant that drafts replies grounded in their own content, and only when it can point to a source.
the challenge
The team had tried an off-the-shelf chatbot that answered confidently and often wrongly. Agents stopped trusting it and turned it off. The problem was never the model - it was that nothing tied an answer back to a real document, and nothing measured whether the answer was any good.
what we did
- 01Rebuilt retrieval over their help centre and internal runbooks, chunked so a citation points to a real paragraph.
- 02Made the assistant refuse rather than guess when retrieval came back weak, and surface that state to the agent.
- 03Wrote an evaluation set from historical tickets so every prompt or model change is scored before it ships.
- 04Put the draft inside the existing tool as a suggestion the agent edits, not an auto-send.
the outcome
- [METRIC] of tickets now start from an assistant draft instead of a blank box.
- First-response time dropped by [METRIC] over [TIMEFRAME].
- Agents report the drafts are worth editing rather than deleting - the reason the first attempt failed.
Bracketed values are placeholders until this study is cleared to publish.