There's a group of companies that consistently falls between two stools in the AI conversation. Too big for the stories about the sole trader automating their invoices, too small for the enterprise stories about Microsoft, SAP and Oracle. Companies with ten to a few hundred people, who have been investing seriously in software for years.

In the Netherlands, that group has a recognisable characteristic: they run on AFAS, on Exact, or on one of Visma's brands. And that changes the question. Not "should I do something with AI", but "what do I actually still need on top of this".

01 / the starting positionWhat makes this group different?

That the most important decision has already been made. There's a system in place where the administration, invoices, hours, often also salaries and customer data are recorded in a structured way. That's exactly what most companies without an ERP don't have, and it's the hardest layer to build afterwards.

The scale is bigger than you might expect. Exact states that Exact Online works for more than 675,000 companies. AFAS reports more than 14,000 organisations and 3.8 million employees who receive their payslip through that software each month. Visma isn't a single system but a group that has acquired Dutch brands, including AccountView, Raet and Yuki.

The question isn't whether you need a system. You already have one. The question is what falls outside it.

A caveat about the term itself: "SME+" is not an official category. The CBS doesn't recognise it and there's no threshold above which you fall into it. It's market jargon for companies that have already taken the step towards proper systems. Useful as a description, unsuitable as evidence.

02 / what you already getDo AFAS and Exact build AI themselves?

Yes, both of them, and that's the most important fact in this entire piece. Anyone who brings in an external party without first checking what's already in their own licence risks paying twice for the same thing.

AFAS bundles its AI functions under the name Jonas. It fills in fields in workflows, converts speech to text, translates documents, analyses purchase quotes and matches CVs with vacancies. Whatever AI has filled in gets a recognisable icon, and an organisation can decide per function whether Jonas is active there.

Exact has the Exact AI Assistant in Exact Online. It helps with navigation, flags anomalies in the bookkeeping and outstanding invoices, answers how-do-I-do-that questions and retrieves information to prepare for a conversation.

Read those two lists slowly. If your specific gripe is in there, the honest conclusion is that you need to talk to your supplier and not to us. That's cheaper, it's better supported, and it won't break with the next update.

03 / the boundaryWhere does your ERP stop?

At everything that never gets entered. An ERP records what is finished: the invoice that's been sent, the order that's been placed, the contract that's been signed, the hours that have been logged. That's not a shortcoming but a design choice. The system is a register.

What's not in it is the work around it. The phone call where a customer explained why they were hesitant. The reason this customer got that discount and the next one didn't. The way you build a quote, which lives in the heads of two people. The email exchange where the agreement was actually made. An AI assistant in your ERP can't get to any of that, because it's not in there.

Rows 2 and 3 are the only place where an extra layer adds anything. Anyone who buys an AI system to unlock row 1 better is, in effect, buying reporting. That's a legitimate wish, but then you're looking for BI tools or a consultant who sets up your ERP better, and that's a different conversation at a different price.

How to work out which source leaks the most for you is set out in your tools don't talk to each other.

Turn the question around and it becomes useful: which work in my week doesn't end up in a record? That produces four recognisable situations. They overlap for most companies, but one or two of them dominate, and that's the place to start.

• Knowledge that never makes it into the system. How you build up a price, why that one customer got an exception. The tell: a new employee can operate all the screens and still can't produce a quote that's right.
• Work that happens before the record. A quote only becomes a record once it's sent. What works here is preparation; what doesn't work here is sending without human involvement.
• Work that runs across multiple systems. Mailbox, calendar, project environment, documents, with a person as the link. Ask, for each piece of data, which system is the source, and which are only allowed to read.
• Exceptions that nobody records. Go through the last twenty jobs and mark what went differently from standard. Whatever shows up two or three times isn't an exception anymore but an unwritten rule.

Keep it small after that. One process at a time, from start to finish. One direction first: start with reading from the ERP, because writing back is a separate decision with its own risk. And one owner internally, who it belongs to after delivery.

04 / the technologyCan a layer like this even reach my ERP data?

Yes, through the official integrations, and those are deliberately limited. That's good news and, at the same time, the place where projects fall apart.

With AFAS this runs via Get and UpdateConnectors. You create an App connector with a token, and you set yourself which connectors that token is allowed to use. That means the boundary of what an external party can see and change sits with you, not with that party's supplier. Always ask for that list before you agree to anything.

With Exact this runs via the REST API, with an App Centre where external apps are published. There's a practical brake on it: a limit applies per administration per minute, sixty calls at the time of writing, plus a daily limit that Exact has adjusted over the years. Check the current limits before anyone builds anything for you.

Why that matters: a system that checks your bookkeeping live for every question hits that limit as soon as more than a handful of people use it. The solution is a copy that updates periodically, and with that you have two versions of the truth. There then needs to be an agreement about which one leads and how often it syncs. Anyone who doesn't make that explicit before building runs into it after delivery.

05 / the riskWhat if my supplier does this themselves next year?

Then you've paid for nothing, and that risk is real. Jonas and the Exact AI Assistant didn't exist a few years ago. Both suppliers have hundreds of developers and a direct interest in pulling functionality in-house.

The rule of thumb that follows: the closer a solution sits to the core of the ERP, the greater the chance the supplier catches up with it. Smarter invoice recognition, better search, flagging anomalies. That's their territory, and there the party that owns the data wins in the end.

What gets caught up on more slowly is the part that's about your company rather than the software: how you write, what you agree, why customers do or don't sign. No supplier builds that, because it's different for every customer. If you're going to buy something, buy on that side. More on where that boundary lies is in what an AI system doesn't do for you.

06 / the testHow do you decide this without a quote?

With four questions you can answer yourself in an afternoon. They work against any supplier, including us.

• Is the answer already in my ERP? If so, call your supplier or implementation partner. Start there and nowhere else.
• Is my gripe on the list for Jonas or the Exact Assistant? If so, it may already be in your licence. Check before you compare anything.
• Which connectors or endpoints does an external party need? A concrete list. Anyone who can't give you that doesn't yet know what they're going to build.
• Where will the copy be stored and who is the processor? If you copy your ERP data to an external environment, you need a data processing agreement. That's not a formality but a legal requirement.

Two more things we'd rather say upfront than afterwards. If your offering isn't settled yet, or your way of working changes every quarter, you're locking down a moving target and waiting is cheaper. And if there's no one internally who becomes owner of the integration, it becomes nobody's job after delivery and it fizzles out within six months. That's not down to the technology.

What else you should be asking any supplier is set out in the questions entrepreneurs ask before they buy an AI system, and the ownership question is worked out separately in whose is it when it works.

This text was produced with AI assistance and checked and approved by a human before publication.