Nordea's AI chatbots handle about three quarters of chats and the bank cites a 30% fraud-loss cut, in Nordea's own numbers

The problem

Nordea runs customer chat “chat-first”: in Sweden, a Nordea application owner says, “if you want to chat with us, you will always chat with Nova first”, and a human agent comes in only on escalation [source]. At that scale the question for any bank is what share of chats the bot actually closes, and whether AI has moved the cost line that matters most to customers, fraud losses ([source]). This case file was opened on the claim that Nordea’s annual report states those outcomes; it does not, so every figure below is traced to the document that actually carries it [source].

What was built

Nova is Nordea’s first-line chatbot, built on boost.ai’s conversational AI platform; boost.ai says Nordea Norway’s team “began building the AI Agent in July 2017 and three months later Nova was live” [source]. A 2022 trade report says Nordea by then used 12 chatbots “powered by natural language processing (NLP) and natural language understanding (NLU) algorithms” ([source]). Nordea’s 2025 annual report describes AI being used “to power virtual assistants that support customers” and says “AI systems also help predict customer churn, detect fraud and accelerate software development”, without naming the fraud system or its vendor [source].

The outcome

The strongest figure is in Nordea’s own investor material. The CIO presentation from its Capital Markets Day of 5 November 2025 states “5.2 million chats; ~78% handled by Nova Chatbot”, with a footnote giving the window as October 2024 to September 2025 [source]. Six weeks earlier, Malthe Falck, Nordea’s Head of Retail Digital, said in an interview Nordea published that “76% of all chat sessions are now handled by our chatbots”, with no period attached [source]. A conference session abstract for Nordea at the Digital Tech Summit 2026 words it as “about 75% of customer interactions” ([source]). The three wordings are kept apart here rather than averaged, and none of them defines “handled”, so this file does not read them as “resolved without a human” [source].

On fraud, the same interview says “we have seen a 30% reduction in fraud losses”, listed among the results of using AI [source]. Nordea published no period, base amount or method for that number, and the Digital Tech Summit abstract says “cutting fraud by 25%” instead ([source]). The annual report confirms the direction but not the figure or the cause: it says Nordea has reduced fraud losses for customers “by implementing safer and easier-to-use security solutions, improving our monitoring, and creating products that are even more secure” [source].

The figure this case was opened on is the weakest. boost.ai’s case study says that in Nordea Sweden “the AI Agent has a 91% in-scope resolution rate for private banking customers and 95% for corporate customers” [source]. The same page defines in-scope resolution as “conversations that are marked as ‘solved’”, so chats outside the bot’s topics are not in the denominator ([source]). This is the vendor narrating its own client, the page is undated, and the Nordea staff it quotes speak only qualitatively [source]. FinTech Futures repeated the numbers in November 2022 as “a 91% resolution rate for private banking customers”, dropping both the in-scope qualifier and the Sweden scope ([source]). A full-text search of Nordea’s 2025 annual report found no mention of a chatbot, an AI agent, the 76%, the 91% or the 30% [source].

The one figure the annual report does carry is volume, not outcome: “1.6 billion personalised in-app actionable insights”, footnoted as a yearly figure “estimated based on the activity level in September 2025” [source].

Why this is not green

Every load-bearing number here is Nordea describing itself, in an interview it published and an investor presentation it gave, and the weakest of them, the 30% fraud cut, has no period or base to check it against [source]. The resolution rate is weaker still, because its only source is the vendor ([source]). What would raise this file is an independent measurement: a Nordic financial supervisor’s customer-service or complaints review naming Nordea, or national fraud statistics from the Nordic banking associations that break out Nordea with a period. None was found. Under TIN’s rule, green never rests on a subject’s own figures, and asking Nordea or boost.ai to confirm them would not count.

Related case files

• Klarna’s AI assistant and its walk-back: the other Nordic customer-service chatbot whose share-of-chats figure came from the company itself.
• JPMorgan’s COIN and the 360,000 hours: another large bank’s AI figure repeated for years without anyone outside the bank measuring it.
• Air Canada held liable for its chatbot’s advice: the risk side of a chat “handled” by a bot: the company stays liable for what the bot says.