Conversational AI in banking: assistants, chatbots and voice at 140 banks.

What is conversational AI in banking?

Conversational AI in banking is software that lets a customer bank in natural language, by text or voice: ask, move money, dispute, get guidance and, increasingly, have the assistant complete the task. 68 of the 140 banks profiled here record a customer-facing assistant, chatbot or voice bot, 47 in production; the wider customer-facing AI count, including agent assist and personalisation, is 95.

Three generations are in production at once. The first is the intent-recognition chatbot: the customer's words are mapped to one of a few hundred intents and the answer is a curated response; Bank of America describes Erica as trained by its own data scientists with a library of curated responses, Capital One built Eno's language processing in-house in 2017, and Fifth Third's Jeanie handles 150 intents. The second uses a large language model to understand the question and compose the answer from the bank's own content, with retrieval so that the answer cites something: NatWest's Cora+ on IBM watsonx, BBVA's Blue, BNP Paribas's HelloïZ on Mistral models, Deutsche Bank's Paula. The third is agentic: the assistant completes the task. DBS's Joy became fully agentic for corporate clients in Singapore in July 2026; Schwab's Charley does transfers, payments, trading actions and beneficiary changes; Citizens' CiZi completes certain transactions.

Voice is the channel that changed most in 2025 and 2026. The touch-tone menu is being replaced by a model that understands conversational speech and either resolves the call or routes it: CIBC's Voice Assistant has taken 16 million-plus calls since its 2025 launch, Busey replaced its IVR with Glia Voice, U.S. Bank plans agentic voice self-service on Amazon Connect, TD targets about 50% of incoming US contact-centre calls with a voice agent, Amex is piloting conversational agents to replace its legacy voice response and USAA plans phone routing that understands conversational speech in 2026. The employee-facing half of this, agent assist and call summarisation, is on the AI in banking customer service page; the personalisation engines behind the insights are on AI customer segmentation.

Which banks run named assistants?

36 named assistants are on the record of the banks here, from Bank of America's Erica (2018; 3.6 billion cumulative interactions) and Capital One's Eno (2017) through Wells Fargo's Fargo (2023; past 1 billion interactions) to the 2026 generation: Schwab's Charley, Citi Sky, DBS's agentic Joy, BNP Paribas's HelloïZ on Mistral models and CommBank Companion. Each row below gives the bank's own description, the channel, the status and the source.

Only assistants the bank itself has named and described are listed, with the bank's own description of scale. Where the launch year is on the record it is given; otherwise the date is the earliest sourced event on the bank's timeline for that assistant, usually a first disclosure rather than a go-live. Vendor-built assistants are included when the bank or the vendor names the bank.

Two things stand out. The US assistants with the biggest totals are the oldest ones, built before large language models and now being rebuilt around them: Bank of America's EricaAssist brings the Erica technology to about 18,000 customer-service desktops, and Truist added a universal “Ask Truist Assist” search in front of its assistant. And the agentic assistants are arriving first outside the US consumer bank, where the rules on automated decisions are more settled: DBS Joy for corporate clients in Singapore, Citi Sky for wealth clients, Schwab's Charley for brokerage customers, NatWest's agentic spending insights as a pilot due to reach 25,000 customers. The full list of every bank with a conversational record, including the ones without a named assistant, is below; 61 rows are in production.

What can the assistants do?

Ten capabilities recur: balances and transaction questions, payments and transfers, spending insights and guidance, dispute collection, voice, more than one language, generative answers over the bank's own content, completing tasks, running inside a third-party assistant such as ChatGPT, and handing off to a person with a summary. The matrix below names the bank whose record shows each.

The matrix lists each capability with the banks whose record describes it; a bank missing from a row has not described that capability in public, which is not the same as not having it. The bank page carries the source for each entry.

The capability that separates 2026 from 2023 is the last four rows. A generative assistant that answers from the bank's documents has to be grounded and checked, which is why NatWest describes governance guardrails on transparency and trust around Cora+ and why Fulton warns customers that its chatbot may not always have the right answer. An assistant that acts has to be scoped, which is why DBS started Joy's agentic servicing with corporate clients and plans card-blocking for retail later. And an assistant inside ChatGPT, as NatWest and BBVA have built, moves the conversation onto a platform the bank does not control; both limit it to information and public APIs. The AI agents in banking page covers the controls.

What scale do banks report?

Bank of America's Erica: 3.6 billion cumulative interactions, 24 million-plus active users, 2 million a day. Wells Fargo's Fargo: more than 1 billion interactions in under three years, over 160 million with Spanish-speaking customers. NatWest's Cora: 12.9 million conversations in 2025, half with no human intervention. Garanti BBVA's Ugi: about 6.4 million interactions a month. ING: 75% of chats resolved without a person in 2025. Commonwealth Bank: 86% of conversations resolved end to end. Citizens: call volume down 44% with CiZi. KeyBank: about $0.25 per AI-handled call against about $9.

The banks' own figures as reported, with date and source. Interaction counts, user counts, containment rates and cost per contact measure different things and are not comparable across banks; a containment or resolution rate depends on which contacts are routed to the assistant in the first place. Vendor-reported figures are labelled.

The resolution series are the most useful numbers because they show the trend inside one bank: ING's chats resolved without a person went 46%, 54%, 75% across 2023 to 2025 as generative AI took over routing; Société Générale's BoursoBank went from 39% of client interactions handled by AI in 2020 to 65% in 2025; NatWest's Cora went from 50% of conversations with no human in 2025 to 3.8 million of 7.1 million fulfilled entirely digitally in H1 2026 after generative journeys grew from four to 21. The demand side is in one survey: TD found 69% of US consumers would let a bank AI assistant complete a task, and 83% of those who had used AI for a financial decision verified the answer.

What do the CFPB and the states require of bank chatbots?

The CFPB's 2023 spotlight holds the bank responsible for the chatbot's answers, requires it to recognise disputes and to give access to a human, and warns of doom loops, impersonation and data exposure; its 2024 comment to Treasury says there is no new-technology carve-out. From January 1, 2027 Colorado's HB 26-1263 and California's AB 1609 require disclosure that the customer is talking to AI and, in California, a good-faith route to a person. In the EU, Article 50 of the AI Act requires the same disclosure; the UK's Consumer Duty and Hong Kong's generative-AI circular add outcome and fairness duties. A vendor-built assistant is a third-party relationship under SR 23-4.

No US banking rule is written for chatbots; the expectations come from consumer law applied to the channel, from third-party risk guidance when the assistant is bought, and since 2026 from state disclosure laws. The table lists the tracked documents with what each asks of a bank; the AI chatbots in banking page adds the timeline and a risk table, and the by-use-case page lists every chatbot-related document.

The practical consequences on the record: Enterprise Bank & Trust's virtual assistant gives no financial advice and hands off to a client service officer; Banner publishes guidance on how to talk to its assistant and keeps staff hours alongside it; PNC says it is deliberate about how its planned assistant will be labelled; First Interstate published guidance for customers on inaccurate AI-generated rate summaries from third-party search engines. For EU and UK banks the disclosure duty already applies through the AI Act and the Consumer Duty, so the generative assistants there (BBVA, BNP Paribas, NatWest, Lloyds) were built under it.

Which vendors build the assistants?

7 vendors in this site's directory sell conversational AI to banks, led by Personetics (5 banks), Glia (4 banks), Talkdesk (3 banks). The largest banks build on Google Cloud, IBM watsonx, Mistral, OpenAI and Amazon Connect; regional and community banks buy from Glia, Talkdesk, kore.ai, LivePerson, Cisco and NICE.

The table lists the vendors in this site's conversational AI category with the number of banks in the US set that name them on the record. The platforms the largest banks build on are not in the category because they are model and cloud vendors: Google Cloud (Fargo), IBM watsonx (Cora+), Mistral (HelloïZ), OpenAI (NatWest's spending-insights pilot), Amazon Connect, Lex and Bedrock (U.S. Bank, WaFd) and TD's in-house Layer 6. Contact-centre platforms with built-in assistants (Cisco Webex at Columbia, NICE at Fulton, LivePerson at Frost, kore.ai at OceanFirst) are on the customer service page.