Cost invoice workflow: the numbers for an accounting firm
Take an accounting firm serving around 120 clients, where 4,200 cost documents pass through the team every month: purchase invoices in PDF, scans, email attachments and files from supplier portals. Someone has to open each one, read it, match it to a supplier, code it against that client's chart of accounts and enter it. Below we break that down into numbers. The usual caveat up front: this is a worked example built on typical values, not a report from a specific implementation — so that you can check it and drop in your own figures.
- This is a worked example built on typical values, not a write-up of one particular firm.
- At 4,200 documents at 3.5 minutes each, the manual workflow costs about PLN 9,310 a month.
- An agent taking on 68% of documents leaves the team 78 h a month instead of 245 h.
Starting point: 4,200 documents a month
We count the same three values as always: the number of cases, the average time per case, and the hourly rate. Here: 4,200 documents, 3.5 minutes per document, PLN 38/h. That is 245 hours and PLN 9,310 a month. Worth noting that reading the data off the invoice is the smaller part of that time — most of it goes on matching the supplier, working out the right account, and deciding whether the document belongs to that client at all.
What the agent does with a single document
The agent works on the same mailbox and the same accounting system the team uses today. One document goes through four steps:
Reading the document. Pulling the data out of a PDF or a scan: seller, number, dates, amounts, VAT rates, line items. Poor-quality scans go into a separate queue rather than into guesswork.
Matching the supplier and the client. Linking to the supplier records by tax ID, allowing for the fact that the same company often sits in the database under several names. No unambiguous match means escalation, not picking the nearest one.
Coding against the client's rules. Assigning an account based on how this supplier has been coded for this client before, plus rules agreed with the accountant. For a new type of cost the agent proposes but does not decide.
Entry or approval queue. A standard document goes into the system with a full description. A doubtful one lands in a queue with a ready proposal and a note on what raised the doubt.
The team's monthly time spent on the invoice workflow before and after implementation, at 4,200 documents and 68% handled without a human. Includes the time spent checking escalated documents.
What changes in the numbers
Assume the agent handles 68% of documents on its own. The team is left with 1,344 documents a month, roughly 78 hours instead of 245 — 167 hours go back into the work there is never time for in peak season. The cost of handling one document falls from PLN 2.22 to about PLN 0.81, and the monthly cost of the process from PLN 9,310 to about PLN 3,400 including the cost of the system. Over a year that is roughly PLN 71,000 from one process.
Beyond the numbers, the shape of the month changes too. Manual processing piles up before deadlines, so the first days of the month usually mean overtime and the last ones idle capacity. The agent works steadily as documents arrive, which flattens the peak more than the hours saved alone would suggest.
Where this calculation does not hold
Two places. First: the quality of the incoming documents. With a large share of poor scans and photos taken on a phone, the automated share drops well below 68% and 45–55% is the more realistic figure. Second: the number of clients with unusual charts of accounts. Each one is a separate set of coding rules, and with a lot of small clients on individual rules the effort of setting those rules grows faster than the saving.
There is also a line we do not cross whatever the numbers say: the agent does not decide tax treatment in doubtful cases. It prepares the document and a proposal; the accountant makes the call. Professional liability is not something you can hand to a system, and no savings calculation changes that.
Frequently asked questions
How is this different from plain OCR?
OCR reads the content of the document and stops there. The biggest part of the time in this process is not reading, it is matching the supplier and coding against a specific client's rules — the work that comes after reading. The agent does that second step and escalates when it has no basis for a decision.
Do we have to change the accounting system?
No. The agent connects to the existing system through an API, the database or exchange files — including older programs with no modern interface. The team works in the same tools as before.
What about documents the agent does not recognise?
They go into a queue with a note on exactly what the problem was: an unreadable scan, no tax ID match, a new type of cost. The accountant decides, and the decision feeds back into the rules, so the same case usually needs no decision next time.