Accounting with AI: what does the software really automate?

The four levels of administrative automation

To understand what software can really do, it helps to distinguish between four technological levels that are often mixed up in practice:

1. Rule-based automation: These are fixed 'if this, then that' rules. When an incoming payment comes from a known IBAN and contains a fixed description, the software links it straight to a fixed counter-account. This is not AI, but predictable, deterministic code.
2. Optical character recognition (OCR) and data extraction: The software reads a scanned image or PDF file and recognises fields such as invoice date, invoice number, total amount and VAT rate. Modern models can interpret layouts, but require clear scans to assign fields correctly.
3. Predictive booking suggestions: An algorithm analyses earlier entries in your administration and, based on historical patterns, proposes a general ledger account or cost centre. As a user you must check and approve this proposal.
4. Interactive AI assistants: Tools that answer user questions based on language models, search documentation or generate administrative summaries based on typed prompts.

For a complete picture of accounting packages, also see our overview of accounting software.

What documented suppliers actually offer

Different software suppliers apply AI components in different ways according to the sources checked as of 10 September 2026:

• AFAS SB (Assistant Jonas): According to the official documentation on the AI assistant Jonas at AFAS SB, this assistant answers user questions based on help information and product knowledge. Jonas explains how to apply settings and where to find functions. This supports the help function. It does not prove which bookings the package can also process automatically. Please note: Jonas within AFAS SB differs functionally from the capabilities within the larger ERP platform AFAS Profit. Also see our AFAS SB overview.
• Boekie: According to the product page of Boekie AI, this software works as an AI integration on top of Exact Online. Boekie invoices per four-week period and is not a complete standalone accounting package. Commercial claims about time savings or accuracy come from the supplier. See our Boekie software profile.
• Exact Online: Exact describes online bookkeeping and automation. For AI, ask about the specific function, edition and demonstration: which data is proposed, which checks are set up and what happens without user action?
• e-Boekhouden.nl: the supplier gives an explanation of AI and bookkeeping. Such an explanation is not yet a specification of your subscription. Look at the software profile and have the automatic processing you want demonstrated with an example.

Four exceptions for a demonstration

Use fictional documents to assess how the software deals with doubt. You determine the expected outcome in advance with whoever checks the administration. That way you assess the system on its processing and not on a plausible-sounding explanation.

An invoice with two lines

Create a fictional invoice with €100 and €50 in costs that belong to different cost centres. The total is €150 before any VAT. Check whether both lines are recognisable and whether you can choose the cost centres separately. A correct total does not prove that the split is right.

An unclear supplier name

Use two fictional suppliers with almost the same name. Let the system make a suggestion without using a real bank account number. What information determines the match? Low confidence should be visible before the wrong supplier ends up in the administration.

A credit note of €200

Create a credit note for a fictional invoice of €1.000. After correct processing, €800 remains payable, as long as no payment has been made yet. Check the sign, the link with the original invoice and the outstanding item.

The same document twice

Submit a copy of the same invoice, with a different file name. See whether the system flags a possible duplicate. Then send a genuinely different invoice from the same supplier with the same total amount. That must not be discarded as a duplicate without review.

For each example, note the suggestion, the correction needed and what is in the administration after saving. A usable AI function makes uncertainty and recovery visible. A convincing answer in a chat window alone is not enough.

Practical examples of prompts and data protection

When using AI assistants, users ask targeted questions in natural language. A few representative examples:

Example prompt 1 (fictional context):
"Show an overview of all outstanding purchase invoices with a due date within seven days, sorted by invoice amount."

Example prompt 2 (fictional context):
"Show the help instructions for creating a deviating payment term for foreign debtors in this system."

Use fictional data in a first trial. Before real use, discuss which data the assistant can access, where input is processed and stored, and whether existing access rights are followed. Also ask whether conversations are used to improve models. Make the answers fit your organisation's privacy and information policy.

Selection criteria for AI features in administration software

Use the table below during an evaluation to assess what an AI function contributes and which verification steps remain necessary:

Agree on checking and recovery

Agree who approves suggestions, who reviews exceptions and how a wrong booking is reversed. Periodically check a sample of accepted suggestions. Include new suppliers and deviating documents in it. That shows you whether the automation still fits changes in your administration.

Compare Boekie as an addition to an existing environment with the AI features of a full accounting package. Include the costs of the underlying software. You are looking for a workable administration; an AI assistant is one part of that choice.