ING's generative-AI chatbot: chats handled without a human rose from 46% to 75%, in ING's own numbers

The problem

McKinsey’s account of the starting point is that ING heard from 85,000 customers a week by phone and chat in the Netherlands, that its classic chatbot resolved 40 to 45 percent of those chats, and that this still left 16,500 customers a week needing a live agent [source]. That baseline comes from McKinsey, the firm that built the pilot, not from an ING document, so treat it as the provider’s framing ([source]).

What was built

McKinsey says the generative-AI chatbot was built with its QuantumBlack unit and first released on testing days to 10 percent of real customers using the support chat in ING’s Dutch mobile app, launching in September 2023 [source]. ING’s own engineering blog describes the scaled version as a hybrid: it keeps intent-based NLP and adds GenAI rather than replacing one with the other, supported by “Google’s Conversational AI platform” and retrieval-augmented generation ([source]). ING’s 2025 annual report says its “contact centres underwent a GenAI transformation, with chatbots now live in seven countries (the Netherlands, Belgium, Germany, Spain, Italy, Romania, and Australia)” [source].

The outcome

ING put the before/after in its own investor presentation of 18 March 2026: “chatbot deflection”, which it defines as “the percentage of chats handled by the chatbot without human support, at year end”, was 46% in 2023, 54% in 2024 and 75% in 2025 [source]. The same slide does not say which countries that figure covers, and it does not label the metric as GenAI-only, so it measures ING’s chatbot as a whole ([source]).

On its GenAI slide, the same presentation states “>80% of chats resolved without human support in 4 countries, rolled out in 7 countries” [source]. The four countries are not named and the figure is a floor, so it is kept separate from the 75% group figure rather than reconciled with it ([source]).

A wider measure moves the same way. ING’s annual report says that through “expanded and improved digital services”, “including GenAI chatbots”, it “reduced inbound contacts to contact centres by 43 percent (2024: 26 percent)” [source]. The presentation defines that as contacts handled with human support, call and chat, against a 2021 baseline ([source]). ING credits several digital changes for it, so it is not a chatbot figure [source].

The figure this case was opened on is weaker. McKinsey says that “within the first seven weeks of use” the pilot chatbot was “helping 20 percent more customers avoid long wait times” compared with the previous solution [source]. That is the provider describing its own project: ING’s chief operating officer Marnix van Stiphout and chief analytics officer Bahadir Yilmaz are quoted on the page only in general terms, and no ING filing or ING release found for this file repeats the 20% ([source]).

Why this caps below green

The weakest source carrying weight here is the strongest one available. ING’s investor presentation is the only place the 46%-to-75% series and the 80% figure appear, and it is the company describing its own operation to investors, not a measurement anyone else made [source]. The annual report backs the seven-country rollout and the contact reduction, but it is also ING’s own document, and KPMG’s reasonable assurance covers the financial statements, not these operating figures ([source]). Under TIN’s rule, green never rests on a subject’s own numbers. No independent measurement of ING’s chatbot resolution rate exists on the public record, so the honest ceiling is corroborated.

Related case files

• Klarna’s AI assistant and its 2025 walk-back: the other headline European customer-service chatbot, whose chat-share figures also came from the company itself and later met a quality reversal.
• Danske Bank’s fraud-detection machine learning: another large bank reporting its own AI outcomes from its own disclosures, capped for the same first-party reason.
• Air Canada held liable for its chatbot’s advice: the risk side of “resolved without a human”: a bank or airline stays liable for what its chatbot tells a customer.