This is an illustrative scenario, not a real client. It shows how we'd approach a situation like this. Figures describe what we'd aim to measure, not results we're claiming.
The situation
A thirty-person accountancy practice has years of guidance, templates and email threads spread across shared drives and inboxes. Junior staff ask the same questions repeatedly, senior people lose time answering them, and the answer often already exists somewhere nobody can find.
Where AI could help
The honest opportunity is knowledge retrieval: a way to ask a plain-English question and get an answer grounded in the firm's own documents, with a link to the source. Done properly it cites where each answer came from, so people can trust and verify it.
How we'd approach it
- Intake. Which questions get asked most, and where the trustworthy source material lives.
- Opportunity scoring. We'd be honest about document quality, because retrieval is only as good as what it's searching.
- Proof of concept. We'd build a working demo over a slice of the firm's real, anonymised material and let the team ask it real questions, checking the answers cite the right sources.
- Report. Accuracy, the risk of confident-but-wrong answers, how to keep it current, and the running cost.
- Handoff. A development-ready brief, with the guardrails spelled out.
What success would look like
We'd measure how often the tool returns a correct, well-sourced answer on the firm's real questions, and the time saved versus asking a colleague. If it can't answer reliably with citations, it isn't ready, and we'd say so plainly.