Oh great—my bank is doubling down on AI

Well, great. My bank is going to use even more AI.

That was roughly my reaction when I read that Rabobank will invest up to €2 billion over the next three years in its data and IT foundation, customer experience, and scaling artificial intelligence. Only half sarcastic, to be honest. Not all of that money goes to AI, and for a bank with millions of customers, vast troves of sensitive data, and an IT landscape that needs constant renewal, that kind of investment feels necessary, not excessive.

In fact, by 2026 I’d be more worried about a major bank that isn’t putting serious money into AI and IT. Let AI spot fraud faster, check documents, flag unusual transactions, and help employees retrieve information. Replace systems that probably still remember the guilder. Automate the admin no one enjoys. No need for nostalgia there.

But when AI is tied to the promise of faster and more personal service, that’s when I start itching. Faster is often doable. Cheaper for the bank, too. More personal is another story.

Rabobank paints an appealing picture on its site: AI gives employees faster access to information, freeing up time for meaningful customer conversations. If the bank delivers on that, I’ll be the first to applaud.

But we also know how these ambitions can end inside big organizations. “More time for personal contact” slowly morphs into a target to handle 80% of all inquiries without a human. Human support becomes harder to reach, and the experience still gets labeled “personal” because the chatbot uses your first name, remembers past messages, and perfectly times a note saying it understands how frustrating this must be.

A solid language model can analyze my transaction history, predict my likely question, and craft a response tailored to my situation. That can be genuinely useful. The bot might even sound friendlier than a rushed employee after seven angry calls.

The difference isn’t just tone. It’s responsibility. A chatbot can say it understands my frustration, but it doesn’t face the consequences if my account is wrongly blocked. It can apologize profusely without the authority to fix the error. It can mimic empathy while no one is actually accountable.

At a bank, that’s not a philosophical footnote—trust is basically the entire product. Right?

The easy conclusion is that people simply prefer talking to people and don’t trust algorithms. The research tells a more interesting story. Note: these are select studies—there’s no single answer; every situation and person differs.

In six experiments on “algorithm appreciation”, people sometimes weighed advice more heavily when they believed it came from an algorithm. For objective, analytical, and measurable tasks, a system can seem expert and consistent. Makes sense. For spotting patterns in millions of transactions, I’d also pick a good model over an employee with a highlighter.

That trust shifts when a task gets more subjective. Research on task-dependent algorithm aversion, based on four lab studies and two large field studies, shows people trust algorithms less for tasks they believe require intuition, judgment, or human discernment. The authors rightly note consumers often underestimate what algorithms can do. So it doesn’t prove a person always performs every subjective task better. It does show people draw clear lines between contexts.

The outcome of a decision matters too. In research on consumer reactions to algorithmic decisions—including a credit application experiment—participants reacted less positively to favorable decisions made by an algorithm than to the same decision made by a human. A human approval felt like recognition of the individual applicant; an algorithmic approval felt distant.

A pre-registered Hungarian experiment with 2,100 people reached the same split in 2025. Decisions with human involvement were seen as more trustworthy for medical diagnoses, hiring, and transportation than processes using AI. Only for financial investment decisions did the researchers find no significant difference.

People aren’t simply for or against AI. They seem quite good at sensing when a task mostly needs compute power—and when they need context, recognition, or someone accountable for the decision.

For a bank, it’s tempting to sort customer contact into neat buckets. An address change is simple. A question about a payment is simple. Blocking a card is simple. Put a chatbot in front, count how many conversations automation deflects, and present the percentage at the next board meeting.

For the customer, that tidy separation often doesn’t exist.

A question about an account may be part of dealing with a death. A blocked payment might mean someone abroad can’t access their money. A mortgage question could stem from a divorce. A report of an unknown transaction might be identity fraud. An entrepreneur asking about credit might really be asking whether the business will survive next month.

On the bank’s screen, each of these starts as a ticket. For the person on the other end, it might be the worst day of the year.

That’s why I find “how much contact can AI take over” less interesting than “which conversations should a bank deliberately keep human.” The value of customer service rarely emerges when everything follows the standard process. Trust is built when the process falls short and someone chooses to look beyond the category the system assigned.

Large organizations understandably track how many conversations they prevent, how fast they handle questions, and what percentage of customers don’t need a human. The risk is when those internal efficiency metrics quietly become the definition of good service.

A conversation that doesn’t happen isn’t automatically a solved problem. Sometimes the organization just shifts the work to the customer—who now has to search for information, rephrase the question three times, and ultimately prove to a chatbot that their situation is exceptional enough for a human.

The bank logs a shorter handling time. The customer loses twenty minutes and still has no answer.

If you steer purely by the number of contacts AI resolves alone, human help becomes a failure of the system. Yet in complex, emotional, or financially significant situations, it can be a mark of good design when the technology deliberately steps aside.

The most useful application of AI in customer contact usually isn’t a chatbot wedged between the customer and the bank. Put AI between the agent and all the systems that currently get in the way of a good conversation.

Have AI summarize a file, retrieve relevant terms, organize prior contacts, and flag missing information. Let it prep forms, surface possible next steps, and search twenty internal systems during the call. That gives the agent room to listen, probe, and explain a decision clearly. And yes, AI—like a human—won’t always do this flawlessly.

AI delivers speed; the agent delivers context, discretion, and accountability. The customer may barely notice a model is involved—except they don’t have to repeat the same story for the fourth time.

For routine matters, automation should of course be the first choice. I don’t need an agent to download a document, change an address, or check a payment’s status. I just don’t want to jump through six digital hoops when my problem isn’t routine.

A clear path to a human agent is therefore essential if you lean on AI as your front line in customer service.