The challenge

A repair quote arrives as free text. It says "turbocharger". The client's parts catalogue calls that part something else, under a code nobody outside the office uses. Labour is written differently by everyone — sometimes with an hourly rate, sometimes as one line, sometimes just a figure at the bottom.

The workshop that sent the quote is its own problem. The same company shows up under three spellings and two addresses on a single tax ID. Sometimes that really is one business with two sites. Sometimes it is a typo.

Until those names become the client's own identifiers, nothing downstream works — no ticket, no cost check, no reporting. So a person did it, line by line.

And behind all of it sat the question every buyer of document automation asks first: fine, but what happens when it gets it wrong?

What we built

Matching here is not one model call. It is four checks in a row, and a record has to survive all of them before it is written into the management system.

Controlled values. Fields the management system only accepts from a fixed list are matched against that list. The agent picks a value that exists; it cannot invent one.
The parts catalogue. Free-text descriptions are resolved against the client's own catalogue — "turbocharger" to the internal part name and code, including the wording partner workshops actually use.
Workshop identity. Each workshop is resolved on name, tax ID and address together. One tax ID with several sites stays one company with several sites. Three spellings of the same address collapse into one record.
The correction database. Every correction an operator makes is stored — what the agent proposed, what the human chose — and the agent reads that store back on the next quote.

Layer 4 is the one that changes the arithmetic. Human review here is not a safety disclaimer, it is the learning mechanism: an operator who fixes a match is not patching one record, they are teaching a rule that applies to every quote after it. The store belongs to the client and it only gets bigger.

Two rules keep it honest. A match the system is not confident about is never written — it goes to a person, with the candidates already ranked. And some quotes carry logic that cannot be mapped at all: a local convention, a note in the margin, a price agreed by phone. Those go to human review by design, and they always will.

Everything runs in the client's own cloud account. The catalogue, the master records and the correction store never leave it.

The results

First-time match rate at go-live was ~80%. As the correction store filled up it moved to ~92% — same quotes, same operators, a system that had read what they fixed.

The rest is not a failure rate waiting to be driven to zero. Part of it is unmappable by design and goes to human review on purpose. A wrong record written silently into a management system costs far more than one that waited for someone to look at it.

Number	What it measures
~80% → ~92%	share matched right first time, no human touch
4 layers	checks a record passes before it is written
Every fix	operator corrections stored and read back on the next quote
On purpose	unmappable cases routed to human review

This is also the difference between a custom system and a generic one. A generic tool matches against a generic dictionary. This one matches against the client's catalogue, the client's master records, and the corrections the client's own operators made last month — so the curve goes up with use instead of sitting flat wherever the vendor left it.

What you get

The matching layers wired to your catalogue and your master records, the correction store behind them, and the review queue where uncertain records land with their candidates already ranked. Process audit first, scope in writing, and 30 days of support. The intake side of the same delivery — reading quotes out of PDFs, photos and handwriting — is Document Intake & Data Entry Automation.

Not sure your master data is clean enough to automate against? Start with the free AI Audit, or book a call.