Extraction & matching

Each document is classified (invoice, receipt, credit note, bank statement, tax document, contract) and then 25+ data points are extracted — vendor details, invoice number, dates, net / VAT / gross totals, VAT rate, currency, suggested category, and every line item. Each document gets one overall confidence score: at 70 or above the record is created automatically; below 70 the document is re-read with a stronger model, and if it's still unsure, nothing is booked — the document is stored for your manual review. You can edit any field at any time; corrections feed back into vendor learning for next time.

Document scanning produces ONE overall confidence score per document (0–100). At 70 or above, the scan auto-creates the invoice. Below 70, the document is automatically re-read with a stronger model; if confidence stays low, the document is still stored and searchable — you just create the invoice yourself. Bank-transaction matching has its own separate 0–100 score (90+ links silently). The two scales never mix.

Open the document, click the vendor field, type the correct name (autocomplete suggests existing vendors). Tab to save. Future invoices from the same vendor are matched against your vendor registry by name (fuzzy matching handles small OCR variations), and after a few confirmed corrections the system learns the vendor's patterns.

Every imported bank transaction is scored against your open invoices on four signals: invoice number (40 pts), amount (25), date (20), counterparty (15). 90+ links silently, 70–89 links with a review flag, 50–69 waits for your confirmation in the Matching inbox. Matches without an invoice number are only confirmed silently when the payment is recent (within 30 days) and unambiguous — anything older or ambiguous is queued for your review.

There is no separate 'recurring vendors' feature — and you don't need one. Vendors that bill you regularly are exactly what the matching memory learns fastest: after three confirmed matches a vendor is trusted for your company, its payments surface as top suggestions, and its learned category is pre-filled on every new invoice.

Matching rules run inside the engine — they can assign a category to a transaction or ignore it, always with top precedence (rules beat learned knowledge, learned beats AI). There is no self-serve rule builder in the app yet: for most cases vendor learning does the job automatically; if you need a hard rule, contact support and we set it up for your company.

When the AI extracts a non-base-currency invoice, we convert it to your base currency. If the receipt itself shows the rate you were actually charged at (Stripe, Wise, PayPal, card statements), that rate wins; otherwise we use the European Central Bank reference rate for the invoice date. Both the original and the converted amounts are stored, with the rate and its source — ECB rates are what most EU tax authorities expect, so conversions are defensible.

Every match you confirm or reject — and every category you pick — feeds a per-company memory. Vendors you've confirmed 3 times (with few rejections) become trusted, and their future candidates get a capped confidence boost. Learning can re-rank and propose, but it never silently escalates a match across a tier boundary. Your own rules always win over learned memory, which wins over AI suggestions.

Every match carries an explanation: a plain-language headline, four score bars (invoice number 40 · amount 25 · date 20 · counterparty 15 — a total of 90+ auto-confirms silently), and caveat warnings when something deserves a second look. On mobile, the Matched segment lists every matched transaction; one tap unlinks an auto-match, and an Undo toast brings it back if that was a mistake.

One bank payment can cover several invoices. On web you can split a transaction across invoices manually; when the payment text mentions several invoice numbers that add up to the payment, the inbox shows a grouped "Bundle · N" suggestion you approve in one action. Only same-day, same-vendor, exact-sum bundles ever link automatically — and those are fully reversible. Mobile shows splits but editing them is web-only.

Your documents are encrypted — and still searchable. The search index stores irreversible cryptographic fingerprints of the words in each document (keyed per company), never readable text. Everything the scan extracted is searchable — vendor names, invoice numbers, line items — plus the filename. Search starts from 3 typed characters; on mobile, content hits carry a "Matched in content" label.

Uploading a file that already exists in your company is rejected with a clear duplicate notice that links to the original — before any scanning happens. Bank statements imported twice are skipped whole, so transactions never double. Invoices with the same supplier and invoice number get a soft "potential duplicate" flag. Email-forwarded attachments and re-delivered emails are deduplicated too.

On the web Banking page, open the account's ⋯ menu → Remove account. Removing deletes the account and its transactions permanently; invoices that were paid by those transactions get their payment status recalculated. Tick "Also block this IBAN" to make sure a mis-read or mis-forwarded statement can't silently recreate the account. Blocked IBANs are listed on the Banking page and can be unblocked any time.