Which accounting tasks can AI take on?
AI takes on preparatory work: reading data from receipts, matching payments, flagging anomalies and creating lists. The decision on how to post, whether to pay and what to chase stays with you. The overview shows what AI contributes to each task and what you check afterwards:
| Task | What AI contributes | What you check |
|---|---|---|
| Receipt capture | reads supplier, date, invoice number, amounts and tax rate, and suggests an account assignment | Mandatory details under § 14(4) UStG (German VAT Act), tax rate, date of supply, account |
| Bank statement and open items | matches payments to invoices, including partial payments and bulk transfers | ambiguous matches, early payment discounts, differences |
| Accounts payable checks | finds duplicate invoices, new or changed bank details and price deviations from the order | confirm every changed IBAN via a known contact before paying |
| Dunning | creates dunning proposals from the list of open items and drafts payment reminders | Onset of default under § 286 BGB, interest and flat-rate fee under § 288 BGB, customer relationship |
| Analyses | Monthly overviews, cost centres, deviations from the previous year, explanatory texts | Reconcile totals against the accounts |
Which AI features does accounting software already include?
The common programs read receipts using text recognition, suggest entries or match bank transactions. This is often the quickest way in, because receipt, entry and archive stay in the same system. As of September 2026, the vendors describe their features as follows:
| Program | Features according to the vendor | Focus |
|---|---|---|
| DATEV Unternehmen online | According to DATEV, text recognition identifies over 80 per cent of dates, invoice numbers and amounts and transfers them to the payment order; at the tax practice, automation services generate posting suggestions | Exchanging receipts with the tax practice |
| Lexware Office | Text recognition reads receipt data and links receipt and entry; recognises recurring receipts; checks incoming e-invoices for correctness | Accounting and invoicing in one program |
| sevdesk | AI-assisted receipt capture matches receipts to bank transactions and recommends a posting category; suggests matching payments | Accounting and invoicing in one program |
| Candis | captures, reads, checks and pre-assigns accounts to incoming invoices, checks them under § 14 UStG and manages approvals; interface to DATEV | Incoming invoices and approvals |
The table is not a ranking. It shows that text recognition and posting suggestions are already built into many programs; check these features first. An additional AI agent is worthwhile for tasks spanning several files and systems, such as reconciling a bank export with a list of open items from another program.
What do the GoBD require when AI is involved in accounting?
The same principles as without AI, because they apply regardless of whether a person, a program or an AI makes the posting suggestion. The accounts must give a knowledgeable third party an overview within a reasonable time, and every business transaction must be traceable from origin to completion (§ 145(1) AO, German Fiscal Code). Entries must be made individually, completely, correctly, in a timely manner and in an orderly way, and must not be changed in such a way that the original content can no longer be determined (§ 146(1) and (4) AO, likewise § 239 HGB, German Commercial Code).
How this is to be implemented in electronic systems is set out in the GoBD, a circular from the Federal Ministry of Finance of 28 November 2019. They were last amended on 14 July 2025, mainly because of e-invoicing. For every IT system used for bookkeeping, they require procedural documentation that shows what the system does and what comes out of it; it must match the actual process, as summarised by IHK Munich (Munich Chamber of Commerce and Industry). If you use AI, this step belongs in the description.
- Receipt and entry stay linked. An AI suggestion does not replace a receipt.
- Originals are kept in the archive or document management system. AI tools work on copies and never directly in the archive folder.
- Corrections only in a way that keeps the original content determinable: reverse and re-post instead of overwriting.
- The procedural documentation describes which tool does which preparatory work, who checks and who approves.
- Clarify with your tax adviser whether AI working lists are relevant for taxation. If so, they must be retained for six years as other documents (§ 147(1) no. 5 AO).
- For e-invoices, it is sufficient to retain the structured part; the PDF part of a hybrid invoice only if it contains additional or different information relevant for taxation, such as posting notes (GoBD para. 119 as amended on 14 July 2025). More on the changeover in the article E-invoicing mandate.
- 8 years
- Accounting receipts, since 1 January 2025 (§ 147(3) AO, § 257(4) HGB)
- 10 years
- Books, inventories, annual financial statements, and the work instructions needed to understand them
- 6 years
- Commercial and business letters, other documents relevant for tax
What does a verifiable AI task in accounting look like?
A good task names the input files, the output format, the rule for unclear cases and a control total. Two examples you can recreate with copies of your exports:
Reconcile transactions-2026-08.csv with open-items-2026-08.csv. Create reconciliation-2026-08.csv with posting date, amount, payment reference, matched invoice number and status “unique”, “ambiguous” or “open”. Only match if the amount and the invoice number or customer fit, and do not guess. At the end, give the total for each status.
- Input
- A month's bank transactions as a CSV export and the list of open items from the accounts, both as copies
- Result
- A CSV file with one row per transaction and a total row for each status
- Check
- Reconcile the total of all rows with the month's transaction total; resolve every row marked “ambiguous” or “open” by hand
Check the incoming invoices in the folder invoices-2026-09 against suppliers.xlsx. Create checklist-payables.csv with supplier, invoice number, amount and finding. Findings are: same invoice number from the same supplier, same amount on the same day from the same supplier, IBAN differs from the master data list, mandatory detail missing. Do not change any file in the folder.
- Input
- Incoming invoices as PDFs with a text layer and the supplier master data as an Excel file
- Result
- A checklist as CSV that contains only invoices with a finding
- Check
- Confirm every finding against the original; never clarify a different IBAN by email, but via a known telephone number
If the open items or bank statements only exist as PDFs, you first need a table. How hard that can be is shown by our measurement with a test bank statement: two widely used table recognition tools did not find a single complete transaction row in the two-page statement without table lines. So include the balance check in every task that converts a bank statement: opening balance plus the sum of the transactions equals the closing balance. In the measurement, it was triggered as soon as a single transaction was missing.
What applies to data protection and co-determination?
Bank statements, customer lists and payroll entries contain personal data. If an external AI service processes them for you, it is usually a processor, and you need a data processing agreement under Art. 28 GDPR; the mandatory content is set out in the article DPAs for AI tools. If the provider or one of its sub-processors is located outside the EU, a third-country transfer assessment is added.
- Start sparingly: wage and salary data, sick notes and garnishments do not belong in a pilot.
- No access credentials: PINs, TANs and passwords for online banking have no place in any AI task.
- Check the impact assessment: whether a data protection impact assessment is required is explained in the article Data protection impact assessment for AI.
- Involve the works council if a tool can evaluate employees' behaviour or performance.
Where does AI reach its limits in accounting?
AI makes no tax decisions and bears no liability. Account assignment, input tax deduction, accruals and deferrals, and closing entries remain professional decisions that you or your tax adviser make.
- Photos and scans need text recognition. An AI tool without text recognition only reads PDFs with a text layer.
- Learning suggestions repeat old mistakes. A system that learns from earlier entries also adopts their misallocations.
- Totals do not add up by themselves. Always have a control total output and reconcile it with the receipt or statement.
- Special VAT cases such as the reverse charge or intra-Community supplies need a professional review, even if the AI makes a suggestion.
- Never trigger payments automatically. AI prepares payment proposals; approval follows the four-eyes principle.
How do you introduce AI in accounting step by step?
With one process, one closed month and one control figure, and with the old procedure running in parallel until the results are right.
- Choose a process
For example, reconciling bank transactions with open items. It is frequent, easy to measure and has a clear control total.
- Measure the current state
How long does the reconciliation take today, how many items remain open, and how many errors come to light later?
- Use software features first
Check what your accounting program already matches by itself. Use an additional tool only for the rest.
- Test with copies
Work through a closed month with copies of the exports and compare the result with the actual postings.
- Run in parallel
For one to two months, people and AI do the same reconciliation. Differences show where the task needs refining.
- Document and decide
Update the procedural documentation and decide on the basis of the figures. What a pilot with a clear decision looks like is described in Introducing AI in SMEs.
Reconciling, not posting: Biwak in accounting
Biwak is suitable for reconciliations and lists from files, not for capturing and posting receipts. An accounting task starts with an export: bank transactions, open items or supplier master data from your program, as a copy in a separate working folder. In the desktop app for macOS and Windows, you describe which list should be produced from it; the AI agent reads Excel, CSV and PDFs with a text layer and writes the result as a new CSV file, which you check and, if needed, import into your program. Save older Excel files in .xls format as .xlsx beforehand, because Biwak does not read the old format. The page Convert bank statements shows the route from bank statement to transaction list step by step.
- Reconciliation of bank transactions and open items as a CSV list
- Checklists for incoming invoices: duplicates, different IBAN, missing details
- Transaction list from a bank statement in PDF format with a text layer
- Monthly overviews and explanatory texts from Excel or CSV exports
- no text recognition for photos and scanned receipts
- no connection to DATEV, other accounting programs or online banking; Biwak works with files
- no posting and no payment: the results are lists and drafts
- no archive: the checkpoint only backs up the working folder for one task and does not replace a data backup
Frequently asked questions
May AI post entries independently in accounting?
Suggest, yes; take responsibility, no. Many programs already generate posting suggestions automatically. But the entry must match the receipt, remain traceable and must not be changed in such a way that the original content is lost (§§ 145, 146 AO). Describe the process in the procedural documentation and specify who checks and approves suggestions.
How long do I have to keep receipts that an AI has read?
Accounting receipts for eight years, since 1 January 2025 (§ 147(3) AO, § 257(4) HGB). What counts is the receipt itself, and for e-invoices the structured data record, not the list an AI has created from it. Working lists that are relevant for taxation must be retained for six years as other documents.
Can AI read scanned receipts?
Yes, if the tool has text recognition (OCR); many accounting programs include it for receipt photos and scans. AI tools without text recognition only read PDFs with a text layer. Check the amounts, dates and tax rates read out against the receipt before posting.
Does AI replace the accountant or the tax adviser?
No. AI speeds up preparatory work such as capturing, matching and checking, but makes no tax decisions and bears no liability. Account assignment, VAT, closing and tax returns remain professional work. How tax practices themselves work with AI is covered in AI for tax advisers.
Which data does not belong in an AI tool?
Without a data processing agreement, no personal data, which also means no bank statements with names of customers or employees. Regardless of that, never access credentials, PINs or TANs for online banking. Payroll and health data only in systems that are approved for it and covered by contract.
Sources
- § 145 AO: general requirements for bookkeeping and records
- § 146 AO: rules for proper bookkeeping
- § 147 AO: rules for the retention of documents
- § 257 HGB: retention of documents, retention periods
- Art. 97 § 19a EGAO: retention periods, transitional rule
- Federal Ministry of Finance: GoBD, second amendment of 14 July 2025 (PDF)
- IHK Munich: principles of electronic bookkeeping (GoBD)
- § 14 UStG: issuing invoices
- § 286 BGB: default of the debtor
- § 288 BGB: default interest and other damage caused by default
This text is not legal advice. It is the groundwork we had to do for ourselves, with the legal references, so that your lawyer does not have to start from scratch. Where a question depends on your circumstances, the text says so.