99%+ accuracy in automated B2B inquiry classification, zero fine-tuning

Whose work this is. This system was designed and delivered by DX Heroes, the team that builds SmartStaff — as a custom project, before SmartStaff existed as a product. The client is a large Czech industrial enterprise and stays anonymous at their request. The original case study is published at dxheroes.io; this page retells it because SmartStaff grew out of exactly this work.

The challenge

Every business inquiry arrived as an unstructured document: a PDF, a scan, a free-form email. Before anyone could quote, a domain expert had to read it and pull out more than 30 technical parameters by hand — material, dimensions, tolerances, volumes, deadlines. That took up to five hours of expert time a week, and the experts doing it were the same people the quotes were waiting for.

The obvious fix — train a model on historical inquiries — was not available. There was no labelled dataset, no budget for fine-tuning, and no appetite for a system that guesses confidently when it should be asking.

What we built

• Unstructured document processing. PDFs, scans and emails go in as they arrive; nobody retypes anything into a form first.
• A multi-parameter extraction engine. One pass pulls all 30+ technical parameters out of the document, each with the passage it was read from.
• Confidence-based routing. Every extracted value carries a confidence score. High-confidence results flow straight through; anything uncertain stops and waits for a person.
• A self-improving feedback loop. When an expert corrects a value, the correction feeds back — the same mistake does not come back next week.
• A kanban review board. The uncertain cases queue on a board where an expert confirms or fixes them in seconds, instead of reading whole documents.

The results

• 99%+ accuracy across the 30+ parameters, validated against 60 reference inquiries.
• Processing time per inquiry went from hours to minutes.
• Up to 5 expert hours a week returned to the people the quotes were waiting for.
• Zero fine-tuning. No labelled dataset, no training runs — the accuracy comes from the extraction design and the review loop, not from a custom model.

The domain experts responded positively; several described the result as extremely positive.

What this became in SmartStaff

SmartStaff is this delivery turned into a product you can start using in an afternoon:

• Confidence-based routing became the approval gate: every request stops at the proposed steps and waits for a human, and a confidence threshold guards even the work you choose to automate.
• The extraction engine's habit of citing its source became grounding: every step an AI colleague proposes quotes the sentence from your request it answers.
• The feedback loop became rules: correct a colleague once and every later request follows the correction.
• The kanban review board became the approval screen: uncertain work queues for a person, and a yes takes seconds.

The client above needed a custom build. With SmartStaff, the same pattern — take in unstructured work, do the routine, stop for a human wherever certainty ends — is what you get on the free plan.