Exact Online and AFAS run the administration of a large share of Dutch and Belgian companies. Both have added AI to their software, and both can be connected to other systems. So the question many finance and operations managers ask is a practical one: what does the AI in our ERP already do, and when is it worth building something on top?

This article sets out what each vendor's AI does today, where it stops, and how to connect your own AI to either system without putting your books at risk.

What Exact Online's AI does

Exact has been adding AI to its software since 2017. It currently describes three AI agents in its business software, plus a generative assistant:

  • Purchase Agent. Processes purchase invoices automatically, spots deviations and suggests the right general-ledger account. Exact states an accuracy of 98%.
  • Bank Agent. Processes bank transactions and matches payments to purchase and sales invoices and the right ledger account.
  • Support Agent. Answers how-to questions about the software around the clock.
  • Exact AI Assistant. A generative assistant, launched in 2024, that flags anomalies in the books and helps prepare, for example, client visits.

For bookkeeping inside Exact, that covers a large part of the routine work. Check which functions are included in your subscription; that changes over time.

What AFAS's AI (Jonas) does

AFAS groups its AI functions under the name Jonas. Jonas works mainly inside AFAS workflows:

  • Jonas automation analyses the data in a workflow and fills fields automatically. You decide in the workflow editor which fields and with what.
  • Jonas assistant runs a prompt at any chosen step of a workflow and shows the result to the user, who can read it, copy it and ask follow-up questions.
  • Smaller functions, such as recognising a receipt in AFAS Pocket, turning a voice note into a task, and translating fields.

AFAS describes its approach as AI as a tool, not a goal in itself, and says it works in line with the European AI Act.

Where the built-in AI stops

Both vendors' AI is strongest where the work happens entirely inside their own system and matches what they designed it for. It stops being enough when:

  • The work starts outside the ERP. An order in the body of an e-mail, a delivery confirmation in a supplier portal, a price agreement in a PDF contract. The ERP's AI does not read your inbox.
  • Your own rules apply. Customer-specific prices, approval routes that differ per department, how you handle a partial delivery. The vendor built for the average customer, not for your agreements.
  • The process runs across systems. The order comes in by e-mail, the stock sits in a warehouse system, the booking goes into Exact or AFAS, the customer gets a confirmation. The built-in AI sees one piece.
  • Operations, not only finance. The AI in both systems focuses on administration. Order intake, planning and supplier follow-up are largely left to you.

Connecting your own AI: how to do it safely

Both Exact Online and AFAS can be connected to other software through their official interfaces. That is also the route for your own AI solution: it reads from and writes to the ERP through those interfaces, not through the screens.

AFAS itself is clear about the risk. On its customer portal it warns against a recent rise of untested, uncertified AI integrations, such as MCP servers and AI agents presenting themselves as an "AFAS solution", that can change data or expose it. The same caution applies to Exact Online.

What a safe connection looks like:

  1. Official connections only, set up in the ERP's own way (an app registration or connector, with its own access token), never a colleague's login.
  2. Minimal rights. The AI gets access to exactly what the process needs, for example reading customers and articles and creating draft orders, and nothing else.
  3. Drafts before bookings. The AI prepares; a person books. Automatic bookings only below agreed limits, once the numbers show it can be trusted.
  4. A full log of everything the AI read and changed, outside the ERP as well as inside it.
  5. A way to stop and undo. Revoke the token and the connection stops; every change can be traced back.
  6. Where the data goes. Know which AI model reads your administration, where it runs and whether it trains on your data.

Three examples of what a build adds

Orders from e-mail into Exact or AFAS. An AI reads incoming orders from e-mails and PDFs, matches customers and article numbers, applies agreed prices and creates a draft sales order. The order desk approves and handles the exceptions.

Supplier follow-up. An AI checks open purchase orders against supplier confirmations in the inbox, updates expected dates within a tolerance and chases suppliers who have not confirmed.

Month-end preparation. An AI collects the open items that block the close (unmatched bank lines, missing invoices, purchase orders without receipt), prepares a list per owner and follows up. The booking itself stays with finance.

In each case the ERP remains the place where the administration lives. The AI does the reading, matching and chasing around it.

Where to start

First use what your ERP already offers. If Exact's purchase and bank agents or AFAS's Jonas automations cover a process, that is the cheapest route. Then look at the processes where your team still copies information from e-mails, PDFs or other systems into the ERP. That is usually where a build on top pays off first. Having an AI agent built explains what that costs and what to ask, and AI agents for business covers the guardrails in more detail.