AI in banking customer service in 2026: what the record shows, wins and fines

Every bank deploying a chatbot can tell you what share of chats it handles. Ask who else measured that number and the room goes quiet.

That silence is the most useful fact about AI in banking customer service. The public record holds plenty of bank figures and almost no outside measurement of them, except from regulators, who measured something else entirely.

The scale is not in doubt. The CFPB found that “each of the top 10 largest commercial banks have deployed chatbots as a component of their customer service”, and estimated about 37% of the US population interacted with a bank’s chatbot in 2022 [source].

What AI in banking customer service means

AI in banking customer service is software, usually a chatbot or voice agent, that answers a bank customer’s question or completes a routine request before a human agent is involved, and routes the rest to a person.

The metric every bank reports is some version of deflection: the share of conversations that ended without a human. That is the number to read carefully, because the bank defines it, counts it and publishes it.

How does AI help in banking customer service?

AI helps in banking customer service by absorbing the routine volume, and the banks that publish figures say that volume is now most of their chat traffic.

ING told investors its chatbot handled 46% of chats without human support at end-2023 and 75% at end-2025 [source]. The same presentation says more than 80% of chats are resolved without human support in 4 of the 7 countries where its GenAI chatbots run ([source]).

Nordea’s CIO presentation states “5.2 million chats; ~78% handled by Nova Chatbot” for October 2024 to September 2025 [source]. Its vendor, boost.ai, separately claims a 91% “in-scope resolution rate” for private customers in Nordea Sweden, where in-scope excludes chats outside the bot’s topics ([source]).

AI in banking examples, and who measured each one

The examples that circulate share one property: the only measurers are the deploying company or its vendor. TIN holds case files on both ING and Nordea, and both are capped below green for exactly that reason.

ING’s file notes that the operating figures sit outside the auditor’s assurance of the annual report. Nordea’s file found that none of the chatbot figures appear in the bank’s 2025 annual report, and that “handled” is never defined.

Klarna is the example most often quoted, though it is a buy-now-pay-later fintech, not a high-street bank. It said its assistant handled two-thirds of customer-service chats in its first month and did “the equivalent work of 700 full-time agents” [source].

What happens when the deflection rate becomes the goal?

When the deflection rate becomes the goal, quality is what gives, and the clearest public example is the company with the best-known number.

In May 2025 Klarna’s CEO said cost had been “a too predominant evaluation factor” and that “what you end up having is lower quality”, and Klarna began recruiting human agents again [source].

The CFPB described the customer’s side of the same trade. Chatbots can lead people “in continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human customer service representative”, and these “doom loops” usually start when an issue falls outside the bot’s capabilities [source]. A doom loop that ends with the customer giving up still counts as a chat without a human.

What are the risks of AI in banking customer service?

The risks of AI in banking customer service that the public record actually establishes are regulatory, and they land on the bank, not the vendor.

Hungary’s data protection authority fined Budapest Bank HUF 250,000,000 for AI that analysed its recorded customer-service calls, guessed “the emotional state of the client” and ranked callers by likely dissatisfaction [source]. The analysis was “not in itself unlawful”, one law firm noted; the failures were legal basis, transparency and the right to object ([source]).

New York’s financial regulator reached a quieter finding on the Apple Card. It found no unlawful discrimination in the algorithmic underwriting, but “deficiencies in customer service and transparency” [source]. The bank also dropped a six-month waiting period for appeals of credit decisions ([source]). An algorithm the customer cannot question is a customer-service failure before it is a model failure.

The proof

Three TIN case files carry this argument, each checked against its primary document before it went live. Each states its own limits on the page.

• Klarna’s AI assistant and the 2025 walk-back: every 2024 efficiency figure originates with Klarna; the reversal comes from the CEO’s own words in independent press.
• Hungary’s NAIH and the bank’s emotion AI: the authority’s own decision; whether the fine was paid is not on the public record.
• NYDFS and the Apple Card: a regulator’s report that cleared the lending and faulted the service.

For the wider sector picture, see what the measured record shows on agentic AI in banking.

How the records compare

Bottom line

The obvious reading of this record is that AI now runs three quarters of bank customer chat. The record does not support that; it supports that banks say so. Every figure in the top half of that table was produced by a party with a stake in the result, and every figure in the bottom half was produced by someone with the power to fine. If you are buying or building AI for banking customer service, plan against the second kind of number. A deflection rate tells you how often the customer stopped talking to the bot. A regulator’s finding tells you what it cost when they should not have.