0:00On stageThe session opens with questions for the room: what do you use AI for, what works surprisingly well and what really doesn't work?
0:00Okay, good afternoon, everyone. My name is André Kwakernaat. I once founded Twinfield, an online accounting system.
0:13In early 2025 I started again with my son, with Tibata. We thought: we have a good idea.
0:20We'll come back to that later. AI for accountants, around bookkeeping, HR systems and so on.
0:29Today I'm here with Jennifer. We often present together. Today we're doing a free-format presentation.
0:35So no long slides and no boring story, because we need you. Back then I invited him, because he was very forward-thinking.
0:53He was one of the few already doing something with data. Other lecturers found that a bit scary.
1:02I've known Peter for a long time. André too, 20 years ago, when he started Twinfield.
1:10I thought: he does data, finance and IT. Something forward-looking. I want to be part of that.
1:17I'd love to set up trainings with him. Besides being a chartered accountant, I'm also an educationalist.
1:24And now here we are: product development and training, for accountants and finance.
1:33So here we are with you. Let's start. Three and a half questions for you.
1:47What do you use AI for? What works surprisingly well for you? And what really doesn't work?
1:56And what would you really like? We'll write it down and come back to it in what we show you.
2:07A few minutes, in twos or threes. Share your answers. So off you go. Talk it through.
2:17What the room saidIf you don't know where an answer comes from, your trust is limited. Hence the human in the loop: a person who checks it.
2:17Same as the first gentleman said: trust in the data. If you don't know where it comes from, your trust is fairly limited.
2:25If you do know where it comes from, it works well, of course.
2:33And if you show where it comes from, we check it. So you're actually saying: the human in the loop matters?
2:40The human in the loop. A person in the loop. That you know where it comes from and that you check it.
2:45From prototype to LinkedInConnecting to an accounting system is harder than it looks. Exact and Twinfield each work differently.
2:58In early 2025 I thought: we need to do this differently. I set up a company with my son. How great is that.
3:05We started with a number of tests. Friends of mine have a café, a restaurant and a clothing store.
3:12Can we recognise things? What can we do? So we went on quite a tour.
3:17In the end we came back to my old trade: bookkeeping. Sounds boring maybe, but it's a real challenge.
3:25I came up with Twinfield myself, but there's also Exact, Yuki, all kinds of systems.
3:30How are we going to build that? Our first target group was the assistant accountant. Know someone like that?
3:40When the accountant agrees on something, the assistant accountant has to get to work. They're busy, working overtime.
3:46How can we help them? In August we knew for sure: the first prototype works.
3:55Then we set up the companies. And in January, August, September, October, November, December, January...
4:05...I thought: isn't it time we published something? So I thought: okay, let's put it on LinkedIn.
4:15And the whole thing exploded. We have a simple website where you sign up with your email.
4:22In HubSpot, our CRM system, we automated everything. Then you get an invitation.
4:28You can sign up for a 30-minute call. Then we look: is it interesting for us and for you?
4:33Then we either continue or we don't. You say: we're not continuing? Fine. We're continuing? Then what do you want?
4:37And if it doesn't click, we don't chase those people. Sounds strange maybe.
4:41Because I'm convinced they'll all start messing around themselves. I call them bedroom artists.
4:48They think they can connect the whole world through Claude or whatever. But by now I know: connecting an accounting system is no easy job.
4:57For example, a very simple list of open items as of 31-12-2025. Maybe too much detail, but do that in Exact.
5:04Then the API gives you today's open items. Not the ones from back then.
5:09Exact has no opening balance. Twinfield does. Twinfield has other things again. For example, a fixed-asset system with three regimes, where you make three entries.
5:20If you don't know that, you get everything three times. You need to know all those kinds of things.
5:25We've solved all of that. And you have to do it at high speed.
5:30So I sit with our clients' assistant accountants almost every day, online or in person.
5:36That's how we develop the system. I did that at Twinfield too. But back then you still drove to clients.
5:43Back at the office I told the developers: guys, it has to change. Then I had to wait a few days for results.
5:49Now it all happens online. I log it, the guys read it. Two days later they say: André, we have a good idea.
5:56Why don't we do it like this? Truly bizarre. What we did: my knowledge is in the agents.
6:05That's why it knows what a trial balance is. And all that kind of thing.
6:08Agents and contextA main agent calls in agents that each know one system. The context you give, such as your clients and the rules you follow, helps the agents understand what they are looking at.
6:09The trick is: we've now connected all kinds of systems. Exact, Twinfield. The main agent figures out by itself where to go.
6:18Then it calls in another agent, trained in Exact, Twinfield or SharePoint. And then it has all kinds of helpers that know things.
6:29Sounds strange maybe, but you really have to specialise. An accounting integration isn't one integration.
6:36It's several. And what matters, in the background: context.
6:42If you live in the AI world, you live in context. Those agents need to understand the context in which they see something.
6:49A firm with only clients in the medical sector has no receivables. So it's odd if it then comes up with receivables.
6:56That way you can give it all kinds of context. The firm always wants RJ, RJK, 4410. You have to show those rules.
7:05Or: I want the conclusion first, then the explanation, and then
7:10tables. You can pass all of that along. We had to discover all of that.
7:16VAT in 9 checksMaking a pivot table is easy. Many firms forget the checks behind it.
7:16We calculate the VAT. That really amazes me. In the accounting system I press the button and there's the VAT, right?
7:22What do all those firms do? Reconcile the VAT. All entries up to now, minus previous periods' returns: that's the VAT.
7:30What they forget: doing the checks. What do we look at? You can make a nice pivot table in Excel.
7:40But did you check whether previous periods' returns were paid or received? Which entries are in it?
7:47Are there entries outside the period? What about this? What about that?
7:52We actually have 9 points it checks this on. We also have supplementary VAT returns, and so on.
8:01The system works all of that out and processes it. And these are skills.
8:07We have all kinds of skills. Completeness check: are all entries correct? Am I missing periods? What am I missing?
8:15What's odd? Then figure analysis: what do the numbers look like.
8:20And the VAT reconciliation. What's happening now: as of 1 October, they're starting now.
8:25We now have about 31 or 32 contracts. 5 firms do everything through this system from Q3.
8:34So completeness check, figure review, VAT reconciliation. Their clients aren't used to suddenly getting a figure.
8:45Or to things being checked. Because clients do a lot of the bookkeeping themselves, not the accountant.
8:51So how do you know your reconciliation is right? Well, you see here, let me make it bigger: this one's been paid. See?
9:02That way I can go anywhere. It shows you everything: all revenue, suspense accounts.
9:14And in the end: what do you have to pay? Now people ask: just send it on to Nextens or Hix, then I'll file the return there right away.
9:24We do this purely based on feedback from our clients. We build fast and test it straight away.
9:32Look, if I go back to the main menu. Then you see there are beta features.
9:39We can turn those on and off for you. People are already testing: do I have a fiscal unity, how do I group those sets
9:46of books? Or I want to benchmark, one against the rest.
9:50Or see the current-account relationship.
9:53Looking aheadIf you track how much the client borrows during the year and tell them, they don't find out after the fact.
9:54How great is this, I used to teach business studies and economics too.
10:00How great would it have been if I could have learned how this works here, from real data.
10:09That's a very different way of thinking, isn't it? What often happens: they come from college to a firm and hear: you start at
10:21the bottom, with the bookkeeping. And those guys all think: I'm going to make annual accounts, right? That's a real difference in expectations.
10:28But what if you say: we'll put you in a simulation for 3 months first, with real books, and you learn from that.
10:36Because here I can also ask: how do you get there? And which entry do I make for bad debts?
10:43My point is: accountants mostly write history now. Afterwards, when it's already happened.
10:52Take excessive borrowing: if you're above €500,000, you have to settle up. If you track it during the year and inform the client, they know.
11:04If they don't know, they buy a car now. Then they'd better not have bought that car this year.
11:13Or investments, or the work-related costs scheme: the free allowance for your staff. It's all in there.
11:20It calculates that lightning fast. And accountants do nothing with that today. Only history writing.
11:29That's already happened. Truly bizarre. And the funny thing is: you'd think small firms, but big firms also say: André, how do we
11:38fit this into our infrastructure? Because you make it tangible, they suddenly think: shit.
11:46Those accountants have an annual accounts program with a work programme in it. There they have to say everywhere: yes, no, not applicable.
11:54It drives you crazy. I said to Frits, one of the accountants: you spend all day clicking yes, no, not applicable. Why?
12:00Because that system requires it. I said: but what if we generate that document ourselves and put it in the system?
12:06Frits: then I'm okay. But I don't know if the NBA will agree to that.
12:09So I went to the NBA. What do you think? Great idea. Look, here you see all your control questions.
12:19Deed of incorporation, open receivables: does it match your trial balance? Annual bank statements: do they match the bank?
12:27That kind of thing. They spend hours on that. Set it up right and you're done in a moment.
12:37And the nice thing: then we turn it into a PDF. That PDF shows exactly what's approved and what isn't, and which questions remain.
12:53You also see how it was built up and which sources sit behind it. Your complete audit trail. Never seen before.
13:03And this is my daily work.
13:073 developersTibata has 3 developers. The whole system is described in the wiki.
13:07How many developers do you think I have? How many? 3. I asked the guys: if we went back 2 years, how many would we have needed?
13:21Santiago said: as many as 20 people, André. And the funny thing is: the whole system is documented, it's all in the wiki.
13:30In Claude too. When we build something new, it goes in the wiki. Truly bizarre.
13:38Testing used to frustrate me: it doesn't work again, nobody thought of that.
13:43Now I can hardly find a hole in it. That's truly bizarre.
13:48Then it's tempting to say: just update it. No, I'll still add a few tricks.
13:53But it's truly bizarre, and that's what you get to experience. You can run a prompt 5 times, 10 times, see what comes out and tune the whole thing.
14:04We tune all those skills ourselves. No, that's the human in the loop. I have to say: I approve it.
14:09And the accountant has to say: I agree. My son does his own books.
14:14A friend of my son, and all those people around us: all their own books.
14:19Last Saturday I checked 15 sets of books and tracked down the errors.
14:25Then you get a little document like this: you need to change all of this.
14:29People always love suspense accounts. Sorting out items, you name it. I hand the work back, they adjust it, I run it again and check if
14:39it's right. That saves them thousands of euros at the accountant, I think.
14:50Some book salary to their management company and then in bits to their own company. Well, that's a total drama.
14:57Then you have to piece it all back together. And I solve all that in no time.
15:04I'm sitting on the couch, my wife says: why are you laughing? I say: I got another one right.
15:09AI literacy per roleAI literacy is mandatory, and it differs per role and per use.
15:09Look at your opportunities and options, and how to do it carefully and responsibly. Only very recently has a general basic AI literacy training stopped being enough.
15:30Before, the law said: you need a general basic training. Now they say: when are you doing it right, when do you avoid a fine and comply with the AI rules?
15:41Specific per role, per use case. For example: I use it for forecasts, or to check the bookkeeping.
15:52Per role you ensure the right knowledge: how it works, what the risks are and how it adds value.
16:00Then you comply with the AI regulations. So you no longer need to give everyone a general training.
16:05Does AI literacy fit within your existing system? Because it's part of risk management: compliance in order, the frameworks and the responsibilities.
16:17What are the risks and how will we use it? How do we make it powerful and scalable?
16:25Smart, validated tools can help with that, a whole team of AI. Since August this year, the law includes supervision.
16:39If you can't demonstrate AI literacy, that's an important point of attention.
16:44What will you take away?Talk to a colleague about what you'll take back to work tomorrow.
16:45Finally, we'd like to ask you this: discuss it briefly with each other. Did you gain new insights?
16:53What will you take with you into tomorrow? Turn to your neighbour. What will you take away?
André Kwakernaat and Jennifer KuiperFinance Unlocked · 22 September 2026