CasesCustom-made goods brokerage · quoting operating layer

Quotes once priced by digging back through dozens of messages — now the system drafts them

A company with no factory of its own, making and selling through a network of partner factories. Turning one request into a quote meant a person digging back through dozens of messages. We built items, specs and prices into a structure a machine can read, and on top of it the system produces the draft quote — and what to ask back. A person still issues it.

Challenge · What was stuck

Customers speak in outcomes; factories listen in specs. “Logo on the chest” and “about A3 size” sit at one end; dimensions, ink counts and color standards at the other. That translation took up much of the operating conversation. The same item carries different rates and expertise at each factory, so where to send it was a judgment call every time, and context was scattered across request forms, consultations and phone calls — what was agreed on the phone existed in no record at all. The bottleneck was not calculation. It was translation.

Solution · What we built
  1. Built items into a structure a machine reads — item → meaningful dimension → value. Complex specs are flattened rather than nested.
  2. Preserve two values from one measurement — the actual figure production uses, and the band pricing uses. The same fact demands a different language at each layer.
  3. Rules and a dictionary are the source of truth for classification; AI only proposes. The same input must produce the same result whenever it's recomputed — that's what makes it auditable. What AI doesn't know is brought to a person; only what a person approves enters the dictionary — and from then on it's settled without AI.
  4. The system sorts what it can do, what it must ask about, and what it can't. Without enough grounds it produces a question instead of a quote; failing that, it says it can't. Where a factory publishes a rate table, that always takes precedence over statistical estimation.
  5. The system never edits the source of truth directly. It raises changes as proposals in an approval queue, each attached to real order examples. As observations accumulate and deviation falls, “this can be automatic now” surfaces on its own.
Impact · What changed
  • Specs a person confirmed accumulate as a source of truth, not as scattered sheets.
  • Quote lines once picked apart by hand are decomposed into structure, so price is explained item by item, dimension by dimension.
  • The more it's used, the less it depends on AI — what's approved becomes the dictionary, and the dictionary becomes deterministic.
  • A person still issues the quote. Unattended issuing is not in the design.

* A real engagement, anonymized with client consent. We don't publish figures about a client's operating scale.

Get Started

Where is one flow stuck in your company?

In 30 minutes we can find where time leaks out, together.

Book a 30-min diagnostic