Customer's message
Underlined words are the evidence behind job card fields.
Case study · working prototype
Turns a messy customer enquiry into a job card, a draft reply and a quote brief. A person approves every message and sets every price. Nothing is ever sent, and the business and enquiries are fictional.
Enquiries arrive in the customer's own words and often miss something the estimator needs. Someone re-types them, works out what's missing and writes back, sometimes twice, before a quote can start.
need a new kitchen asap how much
To start a quote, the estimator needs the room size, the finish, the colour and whether it's installed. Kez gave none of them, and asked for a price.
2 rounds 1
Hypotheses, each with how it would be measured, mostly from the run log. No savings figures: without real enquiries, any number would be made up.
One quieter benefit: what "complete" means for each product is written down in one settings file. Today it may live in one experienced person's head; with the copilot, a new rep asks the same questions on day one.
It doesn't price, send, or connect to an inbox or CRM. It prepares; a person decides. Those limits are deliberate: they keep the first version safe to trial.
Sent to , which sells . Anything else is out of scope.
Select any job card field to see the exact words it came from.
Every approve or reject is logged, in this browser only.
Tested on hand-written messages. Every number here is read from the saved results, not typed in.
AI reads and drafts, code checks, a person decides.
| Step | By | Why that way |
|---|---|---|
| 1 Classify the message | AI | Reading intent needs language understanding; the cheap model is enough |
| 2 Route it | Code | The AI can't talk its way into replying to an upset customer |
| 3 Extract the job card | AI | Messy wording and contradictions need a stronger model |
| 4 Check every quote | Code | A detail without words from the message is downgraded |
| 5 Find what's missing | Code | The business defines "complete", not the AI |
| 6 Draft the reply | AI | Natural writing is what models are good at |
| 7 Screen for prices and dates | Code | A prompt is a request; this is the guarantee |
| 8 Build the brief | Code | Same format every time |
| 9 Approve or reject | Person | Nothing reaches a customer without a person |