Beyond AI vs. humans: how bunq fully integrates AI as its operating system
How AI-with-people helps users get right outcome fastest, with Bianca Zwart, Chief Strategy Officer at bunq
👋 Hey, Dan here,
Ever since my chat with Bianca Zwart last year – about how bunq moves with such ‘un-Dutch’ speed – I’ve been stuck on one question:
Why can some companies operationalise speed at scale, while others struggle?
Increasingly, the answer is AI. But not just applying AI (everyone’s doing that). AI at the core of everything:
Redesigning entire processes, functions, org structures, and roles from scratch around AI – and coming out the other side even faster.
Most companies are tacking AI on top of what already exists, automating the edges, without touching the core.
Bianca’s “Un-Dutch Playbook“ was my most-read post ever, until it was recently overtaken by Hans Scheffer’s “How to build an AI-first company” (also about operationalising speed with AI).
So, I went back to Bianca to find out what bunq is doing with AI today.
Read it if you want to learn:
Why only optimising your current workflow with AI is a trap
How bunq redesigns entire processes from scratch around user outcomes
Real and detailed examples of what bunq has redesigned with AI: processes, features, roles
The ultimate dogfooding strategy: How bunq’s team stress-tested real-time speech-to-speech translation before shipping it to millions of users
Why Bianca believes the AI model itself will never be anyone’s moat – and what actually will be
Note: This piece is about systems and strategy. Next week, Part 2 will dive into the people side: the hard parts of the transition, why bunq disbanded its AI team, and what they look for when hiring now. Subscribe to get it:
Bianca Zwart is Chief Strategy Officer at bunq, the Dutch neobank founded in 2012 that now serves 22 million users worldwide.
She describes her role as being the “voice of the user” – making sure bunq stays user-centric as it scales.
AI, in her view, is not a threat to that. It’s actually the thing that makes it possible.
Let’s find out how:
1. AI vs. humans is the wrong debate
“Often when you go to panels, or read interviews, you see this debate come up: is it AI versus humans?” Bianca told me.
“But I think that debate is framed the wrong way. On one side of the spectrum, you see companies try to automate everything in pursuit of progress. On the other side, there are companies that keep a human in every step of the process – because it feels safer.”
“But neither approach is necessarily better for the user.”
“So the question we ask at bunq is not humans or AI. It’s what will solve this user’s problem properly and get them to the right outcome faster.”
“In some cases, that means AI can handle it entirely. Other times the situation needs human judgment.”
“Implementing AI isn’t necessarily about removing humans from the equation. It’s more about removing friction from the user experience.”
→ Removing humans is not the goal. Removing friction is.
2. AI as infrastructure
“There’s a difference between using AI as a layer on top and AI as infrastructure.”
Here’s how Bianca likes to think about it:
“Are you using AI like you’re downloading an app on your phone – a nice feature on top? Or is it your operating system – your iOS or Android?
“When AI is just an app, it may answer questions for users, or make an existing process faster, but the company still works the same way.”
“When AI becomes the operating system, it changes the work itself.”
AI has become the infrastructure at bunq.
“AI is connected to all the systems needed to take action. Everything internally and externally. Roles have completely changed. It’s less about one technical integration, more about how the company is designed.”
→ Ask whether your AI is an app on top, or the operating system – that tells you whether you’re just applying AI or completely rebuilding with AI at the core.
3. Redesign everything around user outcomes
Let’s double click on that theme.
Bianca explained that bunq redesigns entire processes from scratch around user outcomes, rather than retrofitting AI onto workflows that existed. Here’s what that means:
“Before AI, you’d have a certain process within a team. Every day, you’d try to make each step a little bit faster or more efficient. That meant improving the workflow, but the workflow itself stayed intact.”
“Now, with AI, you define what good looks like first. Then you use AI to get to that outcome. So you’re not just going faster through an old process or preset checklist.”
She gives transaction monitoring as an example:
Transaction Monitoring with AI
“The outcomes of transaction monitoring we want are: protecting our users, meeting our legal obligations, and staying within our risk appetite. We use AI to find the best way to achieve those goals, rather than teaching it to copy or optimise every step of the old process.”
As Bianca explained, that fundamentally changed people’s roles:
“People are still involved [in transaction monitoring], but it completely changed what they do: they spend less time following repetitive steps, and more time setting the right goals, solving hard problems, defining the boundaries of what our AI can handle, and constantly revisiting what the outcome needs to be.”
I asked Bianca what the outcome of this new way of working has been:
“bunq users are now 95% less likely to fall victim to fraud than the industry benchmark, and have seen an 82% reduction in fraudulent incoming funds compared to other neobanks.”
The team also cut false positives by 4x. This means fewer legitimate customers are wrongly flagged, not just fewer fraudsters slipping through. The system runs on a proprietary ML model trained on millions of transactions, alongside things like mandatory face-matching on new device logins and automatic delays on high-risk payments.
The difference and speed shows up in moments like one bunq’s fraud team described to Sardine, the platform behind their fraud detection.
From the article:
“It was ten at night, and we discovered a fraud ring attack,” said Stephanie Rios, bunq’s Product Owner for Fraud Prevention. “We opened up a laptop, launched the dashboard, and within minutes identified what the fraudsters were doing and deployed a rule to stop the attack in progress. We went from alert to resolution in under 20 minutes” – all without an engineer needing to be involved.
→ Before you automate anything, define what good actually looks like – separately from the current process.
4. Question whether a process should even exist
Once outcomes replace steps, Bianca says it’s time to interrogate the process itself.
She first gives an example of when they didn’t yet get this right at bunq:
“In operations, we were using AI to make processes faster, to go through checklists faster, rather than optimising against a goal. And I was looking at a process through a process lens, rather than through an outcome lens.”
But AI actually helped Bianca here too: “AI is really good at forcing you to question your own assumptions.”
So Bianca recommends asking “why does this exist?” before you ask “how do we speed this up?”
And because capabilities keep shifting because of AI, this becomes a constant job to be done:
“I’ve seen examples of processes that were well-designed and right for a year ago, but not for the capabilities we have now.”
She’s clear that this is difficult, and it’s not just a mindset switch.
“I think as with any habit, you need discipline in the way you work. You need to remind yourself constantly that it’s different, that you can do it differently.”
“What I’ve seen is once people see the outcome of working that way: how much better their work becomes, how much faster – they’re very much inclined to want to work that way. That’s the easiest way to get people thinking that way.”
This also connects back to bunq’s founding principle.
“If I think back to when we started bunq, the whole reason was because we thought things could be done better. We always had this aversion to ‘things have always been done this way.’
“With AI, everyone at bunq is challenged to think that way again. Not ‘how can we make the thing we have faster’ – but ‘why do we have this in the first place? And should we?’”
→ A process that was designed well for a year ago can be wrong today, because the capabilities underneath it have changed.
5. Why bunq isn’t optimising for efficiency
“I think a lot of companies start with efficiency: which task can AI take over, how can we do the same work faster, with fewer people.”
At bunq they ask a more fundamental question:
“If AI had always existed, would we have designed work this way at all? Does it need to exist at all?”
“Efficiency usually follows as a result of that, but it’s not the ambition. Implementing AI well isn’t really an automation project. It’s an operating model project. Maybe efficiency is a result of that, but that’s not the goal at bunq.”
“We’re still hiring people. For us, it’s really about how we can redesign work in a way that’s most effective, most user-centric. And have people focus on the things that have the most value.”
“That’s always been top of mind: how do we keep that DNA? I think if anything, AI brings us back to our core as we scale. bunq was always built around speed, transparency, and listening closely to our users.”
“The risk of growth in any company is that those things slowly disappear under more processes, more handovers, more distance between the user and the people building the product. I think AI helps us reverse that.”
“AI isn’t taking us away from what makes bunq different. It’s how we protect our qualities and DNA as we scale.”
→ Don’t assume the qualities that made your early product good will survive scale by default – name them, and protect them.
6. Meet Finn
At the centre of this shift to AI at bunq is Finn.
“Finn is our own generative AI platform. It sits at the core of how we serve users and how we work as a company.”
Bianca describes bunq’s core users as people who live internationally – move across languages, time zones, currencies, borders. “They might need help in the middle of the night, in another language, about an international payment. Finn allows them to do that wherever they are, whenever they need it.”
But Finn is much broader than support.
“We use Finn everywhere externally, but also internally across coding, bug-finding, building, and improving product experiences. It really connects what users need with what our teams are building.”
“Finn is the bridge between us and users.”
7. What Finn does – and how it makes bunq more human
Here’s what Finn does – three real examples Bianca shared:
“A bunq user from Amsterdam was travelling in the Caribbean and needed cash at 4:00am Amsterdam time. They weren’t sure whether their bunq plan allowed ATM withdrawals there. Within 33 seconds, Finn confirmed they could withdraw cash, identified nearby ATMs for them, and explained which cards could be used.”
“A user noticed overnight charges on their OV public transport card and was worried. Finn immediately reassured them that the transactions had not gone through, explained what had happened, and guided them through the next steps with the transport company.”
“A user preparing for a trip to the US asked how they could avoid card fees. Finn explained how: by opening a USD account on bunq, and enabling AutoSelect. It then walked them through the setup process step-by-step, and later helped them understand which travel insurance benefits were included in their plan.”
None of these examples are that dramatic. But that’s kind of the point – nobody needed to escalate to CX, wait on hold, or repeat themselves three times to different people. They avoided those small moments of friction users often encounter with other companies.
And that, multiplied across millions of users, is what defines whether a product feels fast and human, or slow and corporate.
Real-time speech-to-speech voice translation for support calls
I found this coolest example that Bianca shared. It captures how seriously bunq takes shipping AI into everything, even the high-stakes parts.
“We have AI speech-to-speech translation on our SOS hotline. Let’s say there’s an emergency – you can call through our encrypted line in the app, so we both know we’re talking to the right person.”
“You talk to a human, there’s no AI there. But sometimes you might want to speak in your native tongue, because it’s about money, and you might be stressed. With real-time speech-to-speech, we allow you to do that, across 38 languages.”
Bunq was the first bank in the world to introduce real-time speech-to-speech AI translation with AI in its app, back in 2024.
A human operator is always present; the AI translates in real-time between the caller and the operator.
“Let’s say I speak to a human operator who speaks English, and I speak Dutch. I can just speak Dutch, the operator speaks English, and it gets translated in real-time. It feels like we’re having a conversation in the same language.”
Before it shipped, two employees tested it on themselves for three weeks, with no other channel allowed.
“Before we shipped it, we tested it a lot. Our Head of New Products, Tom, and one of our back-end engineers, Nick, actually spoke to each other for weeks through that hotline, through the speech-to-speech. Tom spoke English, Nick spoke Dutch.”
“They said: if we want to ship this, it needs to be really, really good. So for three weeks, no other communication channels were allowed than this one. That let them fix anything that felt awkward or didn’t work.”
→ For any AI feature with real consequences if it fails, make your own team depend on it before any user does.
8. How AI can help you stay human at scale
This is the part of bunq’s approach that Bianca was most animated about. And it flips the usual story that more scale means a company becomes less human.
“For decades, scale always meant standardisation. The bigger a company became, the more generic the experience became. I think AI is the first technology that allows you to personalise at scale.”
“For us, staying human isn’t about whether a person or AI handles the interaction. It’s about the outcome. Did we solve the user’s problem? Did they know what to do next? How did they feel about it? That’s also how we measure it too, with things like user satisfaction, NPS.”
“We’re not trying to maximise how much is done by AI. We’re trying to maximise how much a user feels helped. Because at the end of the day, that’s what makes any experience feel human. Do I feel understood? Do I feel helped?”
Example - Real-time Product Design
The clearest example Bianca shared shows what “real-time product design” looks like at bunq.
“Before AI, if you looked at product design, a team would start with a feature idea. They’d write requirements, specs, design the screens themselves, then build a prototype, and then test whether users understood it. Now we can do that all in real-time. AI can explore different ways of solving the same problem without people having to design the screen first.”
The real example of how it works now she walked me through:
“Imagine someone is abroad and their card payment fails. We want them to understand in that exact moment what happened, what to do, and how to fix it – without contacting support. AI can create and test several different ways of solving that problem at the same time. One version might explain why the payment failed. Another might show the exact amount the user needs to add, because they need to top up their card. A third might take them straight to the right screen to do so.”
It’s basically real-time AI-powered A/B testing.
“What’s so great about AI is it lets us look at what users do as a consequence. Based on what we showed them, did they complete the payment? Or did they ask Finn or still contact us? Because that was the point we were trying to avoid. Based on that behaviour, we can improve it in real-time. So we don’t have to launch one version, wait weeks, manually analyse, and then start again. We can continue to learn from real users.”
Example - Support
“It can also be communication. For every message we send, or screen we show, we’ve always looked at three things, what we call the support triangle:
Did it answer the user’s question?
Is it clear what they should do next?
How does it make them feel?”
“Before AI, that was relatively static: we’d look at support tickets and feedback with a limited sample, and manually decide what it meant.”
“AI now lets us test those same three questions continuously, across support, app design, basically everything we have. Then we can identify: users keep asking the same question, so apparently it’s not clear enough. That’s actually one of the things that excites me most about AI.”
→ Stop picking a single “best guess” solution and shipping it. AI lets you generate several cheaply, so ship more than one and let real user behaviour decide.
9. Your model is not your moat – and what is
Every company building with AI today is using roughly the same models. So I asked Bianca: where does the edge actually come from, if the technology itself is available to everyone?
“The durable advantage is not the model. Models improve, they become cheaper, they get copied. As with any tech advancement, we’ve seen this before: traditional banks started putting more effort into their apps as more competitors came around. I think the product gap will close, but the mindset gap will widen.”
“I think the real advantage is how quickly a company can turn its capability into a better user experience.”
“Any competitor can copy a feature or use the same AI model. What’s much, much harder is to copy is how a company thinks, how it’s organised, and how quickly it learns from its users.”
“Because at the end of the day, people don’t choose a company because of the model running underneath it. They choose it because of the experience: it feels fast, it feels helpful, it feels easy. So long-term, the advantage isn’t having better AI.”
“It’s building a company that uses AI to understand users, continuously redesign processes around their needs, and improve the experience faster than anyone else.”
→ The model is not your moat. How fast your company learns from real users and turns that into a better experience is.
Thanks, Bianca!
You can find her on LinkedIn and learn more about bunq and their AI Finn.
Have a great week, and keep (re)building (with AI) 🙏
Cheers,
Dan 👋