AI in accounting: what to automate in invoices without the risk
AI in accounting means software that reads invoices, receipts and bank statements, assigns costs to categories and matches payments to documents, while a person approves only the exceptions. You can safely hand it the retyping and the reconciliation, which is work that’s easy to check. Tax treatment, corrections and responsibility for the books stay with your accountant and with you.
In short
- AI in accounting is good at reading documents, assigning costs to categories and projects, and matching bank payments to invoices. Tax decisions and the signature on the return stay with a person.
- Since April 2026, most invoices from Polish companies arrive through KSeF, the national e-invoicing system, as XML files, so OCR is pointless for them. You still need to read receipts, invoices from foreign suppliers and other documents outside KSeF.
- The biggest time saving is not reading the invoice but reconciling: which payment covers which document, and what is missing at month-end.
- Do hard matches such as tax ID, invoice number and amount with rules. Give the AI only what a rule can’t decide, and make it flag what it isn’t sure about.
- It pays off when a few hundred documents go through the company every month, or when your accountant keeps waiting on you for missing paperwork. With a dozen invoices, a good invoicing app is enough.
Contents
- What can AI do in accounting, and what can’t it?
- How does KSeF change invoice reading?
- Five jobs worth automating
- What stays with the accountant and the owner?
- What does it look like at Reefa?
- How much time does it save? A worked example
- When it doesn’t pay off
- How to start: a plan for the first month
- Frequently asked questions
I’m writing as a business owner, not as an accounting firm. At Reefa, my cleaning company in Kraków and Katowice, I built the finance module that prepares documents for our accountant myself. Below is what that means for other companies.
What can AI do in accounting, and what can’t it?
AI is good at repetitive work that’s easy to check, and bad at decisions that need tax law and company context. The table shows the split I use.
| Task | What the AI does | What stays with a person |
|---|---|---|
| Reading a document | Extracts the seller, tax ID, dates, amounts and line items from a scan or PDF | Checks fields flagged as uncertain |
| Describing a cost | Suggests a category, project or site based on earlier entries | Approves new and unusual cases |
| Bank reconciliation | Matches a transfer to an invoice even when the reference is incomplete | Explains payments with no document |
| Month-end gaps | Lists which documents are missing and sends reminders | Gets the document from the supplier |
| Tax treatment | Can suggest, but shouldn’t decide | The accountant decides on VAT, deductible costs and corrections |
The last row matters most. A language model can give a wrong answer in a confident tone, and in tax a mistake costs interest and corrections. So in my system the AI never posts anything on its own. It prepares the data, and every decision has a name on it.
How does KSeF change invoice reading?
KSeF, Poland’s national e-invoicing system, takes away the easiest part of the AI’s job. According to the Polish Ministry of Finance page on KSeF, issuing invoices in KSeF has been mandatory since 1 February 2026 for companies with 2024 sales above PLN 200 million, and since 1 April 2026 for everyone else. Invoices have been received through KSeF since 1 February 2026, and the e-invoice itself is an XML file that follows the official schema and carries a number assigned by the system.
What this means in practice:
- Invoices from Polish companies in KSeF arrive as data, not as an image. There’s nothing to read. You just download and describe them.
- Documents outside KSeF still need reading: receipts, invoices from foreign suppliers (software subscriptions, for example), consumer invoices, which the ministry says are outside mandatory KSeF, and invoices from cash registers, which can be issued without KSeF until the end of 2026.
- Transition period. Until 31 December 2026, invoices can be issued outside KSeF if the monthly sales on them, including VAT, don’t exceed PLN 10,000. Some small suppliers will still send you a PDF.
What to check in the software you already use
Before you order anything new, ask the vendor of your invoicing app three questions. Does it pull cost invoices from KSeF on its own, without a manual export? Can you add a document from outside KSeF, such as a photo of a receipt, and tag it with a category? Does it export data in a format your accountant accepts? If the answer to all three is yes, setting it up properly is often enough. A custom module only starts to make sense when documents come from several places or when you run more than one company.
The upshot: in 2026, don’t buy a system just for invoice OCR. The value moves from “read the invoice” to “describe the cost, match it to a payment and show what’s missing”.
Five jobs worth automating
These jobs have one thing in common: you can check the result in seconds, and a mistake shows up at the latest when the month is reconciled.
- Collecting documents in one place. Invoices from KSeF, PDFs from the inbox, photos of receipts from the team. A robot downloads them, names them and files them by month and supplier.
- Reading documents outside KSeF. The AI extracts the fields and rates its own confidence. Fields below the threshold go to a person; the rest moves on.
- Describing costs. Based on history, the AI suggests a category and cost centre. You or your accountant approve only the suggestions that differ from history.
- Reconciling the bank with documents. Every outgoing transfer should have a cost document, and every incoming payment a sales invoice. This is the most time-consuming part and the one that pays back best.
- A list of gaps before month-end close. Instead of a call from your accountant on the tenth, you get a list of payments without documents on the first.
The order matters. If you start with step 4 on data from steps 1 and 2, every reading error shows up straight away, because the amount won’t match the statement. Reconciliation built this way works as quality control for the AI.
What stays with the accountant and the owner?
Everything that needs a decision stays: whether an expense is tax-deductible, whether VAT can be reclaimed, how to book a correction, what to do with an invoice issued with an error. So do the signature on tax returns and responsibility for the books.
The owner keeps two things no software will do:
- Explaining exceptions. A transfer with no invoice, an advance, a refund. The AI will show that something doesn’t add up, but only you know what the payment was for.
- Chasing suppliers. If a supplier doesn’t send an invoice, the reminder can go out automatically, but a person has the conversation.
The good news: an accountant who receives tidy, reconciled data has more time for the decisions you actually pay for. Before you change anything, agree the data format with them. At Reefa the cash-register export matches our accountant’s template 1:1, because otherwise they’d keep retyping it.
What does it look like at Reefa?
The key lesson: most of the work in my finance module is done by plain rules, and AI only where the data is unstructured. The whole system is described in the Reefa system case study.
Rules: invoices, bank and cash
- Sales invoices sync daily from the invoicing app, from three accounts for three businesses in one system. An invoice is linked to a site first of all by the client’s tax ID.
- Reconciliation follows one rule: every bank transaction needs coverage, so income has an invoice and an expense has a cost document. Once a month, on the first day, the system checks all three businesses and sends me the uncovered items on Telegram, marked red.
- Bank matching finds invoice numbers even when the transfer reference wraps across several statement lines, and it removes duplicates when the same transaction came in both from an MT940 file and from the bank connection.
- I moved the cash register into the system retroactively: 312 entries from 21 closed months, reconciled to the grosz against the documents. During calibration, 14 cash withdrawals on the bank statement matched the manual cash report 1:1.
AI: documents without structure
- Monthly staff documents: the AI detects the document type and month, and a signed cash voucher is matched to its cash entry automatically. The system shows who is missing which months.
- A photo of a document sent to the bot: the AI works out whose document it is, what type and which month, and I confirm with a button.
- Contracts: the AI reads the parties, tax ID, company register number and signatories from the PDF in 13 seconds.
Had I started with AI for everything, reconciliation to the grosz would have been impossible to check. The reverse order, rules as the backbone and AI for the exceptions, means every model mistake shows up in the report.
How much time does it save? A worked example
The company in this example is made up, and the numbers are an estimate. A service company has 250 cost documents and 400 bank transactions a month.
| Task | By hand | After rollout |
|---|---|---|
| Collecting and describing 250 documents | 3 min × 250 = 12.5 h | 0.5 min × 250 = 2 h (checking) |
| Reconciling 400 transactions | 1 min × 400 = 6.7 h | exceptions, about 40 × 2 min = 1.3 h |
| Month-end gaps | 3 h of calls and emails | 0.5 h on the system’s list |
| Total per month | about 22 h | about 4 h |
That’s about 18 hours a month, more than two working days for the person who retypes and hunts for documents today. Put your own numbers into the savings calculator before asking anyone for a quote. If the result is 2–3 hours a month, the rollout won’t pay back.
When it doesn’t pay off
Automating accounting makes no sense when:
- There are few documents. With a dozen invoices a month, an invoicing app with KSeF access and a good working relationship with your accountant are enough.
- Your accounting firm already has its own system. Check first what it accepts automatically. Often sending documents in its format is all you need.
- Your data is a mess. If nobody knows which cost belongs to which project, AI won’t guess. Sort out the process first, for example with a process map.
- You want AI to decide on tax. I don’t do that and I don’t recommend it. Responsibility for that decision lands on you anyway.
Watch out for personal data, too. Invoices from sole traders and staff documents contain personal data, so the model has to run on a business account with a data processing agreement. I described the minimum I insist on in the piece on AI in small and medium businesses.
How to start: a plan for the first month
Start by measuring, not with a tool. A plan that gives you an answer in four weeks:
- Week 1: count documents and transactions over the last three months and the time it takes to handle them. Ask your accountant what they most often wait for.
- Week 2: agree the data format and cost-description rules with your accountant. Write them down, because they’ll become the system’s rules.
- Week 3: run bank-to-document reconciliation on one month and compare the result with what your accountant has.
- Week 4: only now add AI for documents outside KSeF, with a confidence threshold and manual checks for fields below it.
If after the first month the saving is smaller than the calculator showed, stop. For choosing the first process, read where to start with business automation, and for when AI should act on its own, the piece on AI agents.
I build modules like this as part of AI implementation. A rollout costs from PLN 15,000 excl. VAT, and the scope depends on how many document sources and companies are involved. Work out the saving in the calculator first, then book a 30-minute call: I’ll tell you straight whether rules and an invoicing app are enough for your company.