79% of client receipts and invoices post into Merit without a human touch
An Estonian accounting firm serves restaurants, logistics, e-commerce and retail clients. Every receipt photo and supplier invoice used to be typed by hand. Now 42 mailboxes and a Telegram channel feed one pipeline, and only the doubtful documents reach an accountant.
Results
Figures are from the pipeline's own logs: a two-week sample of 1 093 documents (25 July – 8 August 2026), a duplicate audit of 1 741 posted documents, and a 54-document reference set used to test every change.
| Metric | Before | After |
|---|---|---|
| Straight-through posting | 0% — every document typed | 79% (864 of 1 093 documents) |
| Duplicates reaching the books | caught by eye, sometimes missed | 0 of 1 741 posted documents |
| Field accuracy on the reference set | not measured | 574 of 581 fields (98.8%) |
| Client companies served by the pipeline | 36 at launch | 70+ ten weeks later |
| Documents with no automatic path | 100% | 2.9% — routed to an accountant with the reason |
The challenge
- Dozens of inboxes, every format
Each client company has its own mailbox. Documents arrive as PDF invoices, phone photos of till receipts, scans, forwarded email threads and spreadsheets.
- Each client has its own chart of accounts
The same supplier maps to different accounts and VAT codes depending on the client. A generic OCR tool cannot know that.
- Broken supplier data
VAT numbers missing the country prefix, registry codes with typos, and suppliers that changed their legal name — all of which block a clean posting.
- Duplicates and silent losses
The same invoice arrives twice (a forward plus the original); slow attachment downloads create copies. Dropping a document silently is worse than a duplicate.
What we built
42 mailboxes and a Telegram bot feed a single queue. Every attachment gets an identity (mailbox + message + attachment) and an atomic lock, so a retry can never process it twice.
A first pass reads the document; if the totals do not add up, a slower second pass re-reads it and only the fields the first pass got wrong are replaced.
A VAT number missing its prefix is reconstructed from the IBAN, address and language, then checked against the Estonian business register and VIES before posting.
1 974 rules across 188 items decide the account, VAT code and cost centre for a supplier line under a specific client. Rules are edited by the accountants, not by us.
Line totals are reconciled against net, VAT, rounding and returnable-deposit amounts. A document that does not add up goes to review, not to the books.
Anything uncertain is posted to a review channel with the reason attached. Accountants fix or approve there; nothing is dropped.
How it went live
Map the flow
Counted the mailboxes, sampled the document mix per client, and agreed what 'posted correctly' means with the accountants.
late June 2026First production run
Live on the firm's own documents from 13 July. A 54-document reference set became the acceptance test for every later change.
2.5 weeks from start36 → 70+ companies
Clients were added mailbox by mailbox. In August the intake was separated from the processing core so rule changes no longer interrupt intake.
July – September 2026Honest limits
Multi-page invoices photographed page by page can still be under-counted — one real case posted €2 831.71 instead of €3 853.75 and was corrected by hand. Cash-register daily summaries are sometimes classified as the wrong document type, and credit notes that reference an invoice in the same email are still handled manually. All three are on the roadmap; none of them is hidden from the accountants.
Frequently asked questions
Does this replace the accountant?
No. It replaces the typing. Accountants keep the mapping rules, review the exceptions and sign off the month. The 21% of documents that need a decision still get one — from a person, with the reason shown.
Why not use a receipt-scanning app?
Off-the-shelf scanners read a document but do not know that the same supplier belongs to a different account for each of your 70 clients, and they do not repair broken VAT numbers against the register. The mapping rules are the part that makes posting automatic.
How long did it take to go live?
About two and a half weeks from the first meeting to the first production run, then continuous rollout as clients were added. Most firms can expect a similar pilot length for one document stream.
What happens when the pipeline is wrong?
A document that fails validation never reaches the books; it goes to the review channel with the reason. When a wrong posting does slip through, as with the multi-page invoice above, it is visible in Merit with the source PDF attached and is corrected there.
Can this work with Xero, QuickBooks or 1C instead of Merit?
Yes. The intake, reading and validation stay the same; only the posting step changes. See the integration notes on the solution page for what each system accepts through its API.
Similar stream in your business?
Tell us which documents or requests eat the hours. Free audit, fixed pilot price, no commitment.
Get a free automation audit