01 · Dissolve
One data layer
Quotes, jobs, customers and contacts extracted from the job-management API into a single shared layer — the same numbers the dashboard shows.
Stage 1 · Dissolve
The problem
As the MD described it- The answers exist. Getting them takes a week.Who sends us work that converts? Who requests quotes that go nowhere? The data sits in the job system — but every question means someone building an export.
- Dashboards answer the questions you predicted.The moment the MD asks something new in a Monday meeting, it joins a list, and the moment passes.
- Sales effort goes where the noise is.The loudest contacts get the attention; the quietly reliable ones — and the time-wasters — look identical in the inbox.
Before: how a question got answered
Mon 09:15 — MD in the sales meeting: "Which contacts actually convert?"
Mon 11:40 — Ops exports quotes and jobs to a spreadsheet, starts matching them by hand.
Wed 16:00 — Draft answer circulated. Two contacts miscounted; redone.
Fri — Answer arrives. The meeting it was for was Monday.
Stage 4 · Perfect
The result — ask, and it answers from live data
Synthetic replay of real question typesStages 2–3 · Separate & recombine
How it works
The build, in four moves02 · Separate
Governed tools
The AI never free-writes SQL or guesses. It calls named, read-only tools — win rates, contact activity, quote ageing — each one testable on its own.
03 · Recombine
Answers with working
Every reply cites the records behind it and keeps figures reconciled with the dashboard, so "Ask" and "Report" can never disagree.
04 · Perfect
Shipped & scoped
Runs on the client's secured server behind their login, read-only, with the token and access model documented in the handover pack.