Fuzzy payment matching even without a variable symbol, AI invoice extraction with human-in-the-loop, connections to bank APIs and ERP — built and operated by the same architect. Not accounting software with an AI sticker on top. We solve exactly where boxed SaaS hits a wall: partial and merged payments, missing payment references, foreign currencies, enterprise volume, custom ERP.
Standard matching via a payment reference (variable symbol) is handled by any boxed tool without any AI — and it should be, it's the right default. The problem is the remainder that doesn't carry a clean reference. It either ends up in a manual queue, or with a vendor who slapped "AI" on the marketing page but never actually wrote the fuzzy logic.
Pohoda, iDoklad and Fakturoid all match a payment when the reference and exact amount line up. But for companies with less disciplined clients, 10-30% of payment volume fails that condition — partial payments, merged transfers, missing references, foreign currency. No boxed tool handles that remainder except by dumping it into a manual queue.
Fuzzy matching (amount within tolerance, counterparty name, account number, date, payment text) and AI invoice extraction with human-in-the-loop is precise engineering — not "throw AI at it and hope." It needs an audit log, idempotency and reconciliation, not just a prompt over a PDF.
An accounting software vendor gives you a support ticket and a roadmap for next year. For a smaller integration it makes more sense to talk directly to the architect designing the logic — for a large, multi-system program, project management has its place.
We don't sell "AI accounting" as a black box. We build a durable, auditable layer on top of your bank account and ERP — the same architect designs the logic and runs the integration in production.
Amount within tolerance, counterparty name, account number, due date, payment text — combined signals produce a match score. Clear cases match automatically; uncertain ones go to a manual queue with a suggestion, not an empty list.
Data extraction from PDFs and scanned documents, with a human confirming uncertain items. No silently posting a misread amount — just speeding up work you already do by hand.
Bank statements, Business Central and other ERPs — durable, idempotent calls, not fire-and-forget scripts. One payment never gets processed twice, even after a restart or an API outage.
Every match, automatic or manual, is recorded — when it happened, which rule, what score, who confirmed it. For regulated or enterprise accounting that isn't a nice-to-have, it's a condition for passing audit.
You talk to the architect who designed and wrote the logic. The same person who built the KSeF integration handling 40,000+ documents handles your payment matching too — on larger programs this pairs well with project management.
A periodic "does reality match?" check — does the sum of matched payments equal what actually landed in the account? Without this layer, matching looks like it works until the bank API fails mid-statement-download for the first time.
We cut bank contract generation from 2 hours to 3 minutes — a 40x improvement, at ČSOB before SolutionBox existed. We built the same durable logic (idempotency, retry, an audit log on every step) for KSeF, where it processed over 40,000 documents with 100% delivery into a regulated national system. Fuzzy payment matching is the same principle applied to a different problem: combining amount tolerance, counterparty name, account number, due date and payment text where companies with less disciplined clients see 10-30% of payment volume arrive without a clean reference.
We'll look at what percentage of your payments clear via a clean payment reference and how many end up in a manual queue. We'll tell you straight whether boxed accounting is enough, or whether custom fuzzy matching and AI extraction makes sense. We'll reply within 24 hours with a concrete number, not a generic deck.
Request AI accounting automation →