Cutting quote turnaround from three days to twenty minutes

Requests for quotes arrived by email, as free text, as PDFs, sometimes as a photo of a drawing. A sales engineer read each one, looked up the parts in the ERP, checked the customer's price list, built the quote in a spreadsheet and sent it back. Three days on average, longer in busy weeks. Competitors answered in one.

The rules were mostly known. The time went into reading, rekeying and checking. Errors crept in at the same places every time: a wrong unit, an outdated price, a discount applied twice.

During discovery we timed the process step by step. Reading and rekeying were two thirds of the effort; the actual judgement, whether to quote and at what margin, was minutes. That told us what to automate and what to leave human.

We built an intake service that reads incoming requests, uses an AI model to extract the requested parts, quantities and specifications, and matches them to the ERP catalogue. The system then drafts a quote using the customer's current price list and margin rules. Anything below a confidence threshold, or above a value limit, is flagged for review. A sales engineer approves, edits or rejects every quote in a single screen before it goes out. Nothing is sent without a person.

The automation ran alongside the old process for three weeks. We compared its drafts with the hand-built quotes, tuned the extraction and the rules, and only then switched over.

A standard quote now takes about twenty minutes from email to approval, most of it the review. The sales team saves around twelve hours a week, which went straight into following up on open quotes. Pricing errors dropped by roughly ninety percent because the price list is read from the ERP instead of remembered.

The project took eight weeks. The company has since extended the same approach to order confirmations.

Next.js and TypeScript for the review application, PostgreSQL with Prisma, the Claude API for extraction with EU data handling and no training on client data, a scheduled sync with the ERP, all containerised on the client's EU VPS. The client owns the code, the prompts and the evaluation set.