Retail Banking Assistants: Used by Millions, but Outcomes Are Rarely Measured in Public

The use case

In 2025, 20.6 million people used Bank of America's Erica assistant, which the bank reports generated nearly 700 million interactions (operator-reported, 2025)². NatWest's assistant, Cora, handled 12.9 million retail banking conversations in the same year (operator-reported, 2025)⁴. These are the largest public usage figures found for a retail banking assistant. Neither source reports how many of those exchanges ended with the customer's problem solved.

A retail banking assistant is software in a bank's app or website that takes a customer's typed or spoken question, such as how much was spent on food last month or how to add a card to a phone wallet, and then answers it, carries out a simple request, or passes the customer to a person.

What counts as AI here. Only systems that use machine learning to interpret free-form language, including generative AI, which writes new text instead of choosing a prepared answer.

What does not. The US Consumer Financial Protection Bureau (CFPB) describes rule-based chatbots that "use either decision tree logic or a database of keywords to trigger preset, limited responses", and states that "much of the industry at least uses simple rule-based chatbots" (June 2023)¹. Human contact centres, app logins and push alerts are not AI either. Bank of America's headline of about 30 billion digital interactions in 2025 consists of 16.6 billion logins and 13.3 billion alerts, and is not a measure of Erica (operator-reported, 2025)².

Staff-facing tools are out of scope, such as the call summaries and simplified complaint responses NatWest credits with 70,000 hours saved in retail banking (operator-reported, 2025)⁴.

Where it is running

The deployments are not new. Erica launched in 2018¹², Cora has run for nine years⁴, and NatWest announced a generative AI upgrade to Cora in November 2023⁵. The two Bank of America user series use different bases (a calendar-year count and an unstated period) and cannot be compared or added²³.

The evidence covers three English-speaking markets. No deployment in continental Europe, Asia or Latin America was verified from an opened source.

What the evidence shows

No controlled trial, regulator audit or independent evaluation of resolution rates was found. The strongest evidence is official complaint decisions and a regulator report, which show how assistants fail rather than how often.

Usage counts (operator-reported). The figures above come from the banks themselves. Usage counts measure uptake, not whether a customer's problem was solved. An interaction count includes every message, so a customer who rephrases a question three times adds three interactions²³⁴.

A decision against a bank (official ombudsman decision, single case). In September 2023 a NatWest customer with a disability asked the bank's chatbot about adding a QR code to his wallet. The chatbot did not understand, did not log his complaint and offered no route to a person. NatWest explained that the chatbot only responds to specific wording. The Financial Ombudsman Service upheld the complaint and added £140 to the £160 the bank had already paid; the ombudsman said the bank should have referred him to a human agent or a phone service. The provisional decision is dated 2 May 2024; the final decision date is not shown⁶.

A decision for a bank (official ombudsman decision, single case). A Lloyds customer complained that he was automatically connected to a virtual assistant at the start of every chat. The ombudsman did not uphold the complaint, finding that Lloyds is "free to design its systems and operate its chat facility in line with its commercial discretion". Customers could ask for an agent. The decision date is not shown; the events date from 2023⁷.

Regulator-reported harms (June 2023). The CFPB lists "wasted time, feeling stuck and frustrated, receiving inaccurate information, and paying more in junk fees", and quotes consumer complaints from 2020 to 2023 about chatbots stuck in a loop with no way to reach a person. It also cites a 2022 survey, reported by Forbes, in which 80% of consumers who used a chatbot left feeling more frustrated and 78% needed a human (self-reported survey, 2022)¹. These figures are over three years old and no newer regulator-level equivalent was found.

An operator claim that was reversed (August 2025). Commonwealth Bank insisted its voice bot had cut call volumes by 2,000 a week (operator-reported, relayed by The Nightly from the Australian Financial Review), then conceded that calls had risen, with managers using overtime and team leaders pulled onto the phones⁹. The bank called the job cuts an "error" and apologised to the staff concerned⁸. No audited call data was published.

What it takes: cost, integration, governance

Cost. No official per-conversation cost was found. NatWest reports investing £1.2 billion in 2025 across AI, data and technology, which is not specific to the assistant (operator-reported, 2025)⁴. Industry savings figures older than three years were left out.

Integration. These are long programmes. Erica has run since 2018 and Cora for nine years²⁴. NatWest has extended generative AI journey by journey, from four to 21 by the end of 2025⁴. The CFPB notes that the statistical methods behind generative chatbots "are not well-positioned to distinguish between factually correct and incorrect data" (June 2023)¹.

Human handoff. Both ombudsman decisions turn on whether the customer had a route to a person⁶⁷.

Rules. The CFPB states that chatbots "must comply with all applicable federal consumer financial laws, and entities may be liable for violating those laws", while noting that the spotlight does not itself impose obligations (June 2023)¹. The ombudsman wrote that NatWest "must ensure the support it provides is effective, regardless of the channel of support used"⁶. Outside banking, a British Columbia tribunal ruled on 14 February 2024 in Moffatt v. Air Canada that an airline was liable for its chatbot's wrong answer and rejected the argument that the chatbot was a separate legal entity; the award was about CAD 650 plus interest and fees¹⁰. In the EU, law-firm commentary reports that providers of AI systems that interact with individuals must tell them it is an AI system, from 2 August 2026 (July 2026)¹¹. The text of the regulation could not be read in full in this research, so the article should be checked.

Where it falls short

Tested against its design, the case for retail banking assistants is thinner than the usage figures suggest.

Design of the evidence. Every scale figure is an operator-reported count with no control group and no independent audit²³⁴.

Whether gains persist. No sustained gain is documented. The one testable operator claim, fewer calls at Commonwealth Bank, was undone by the bank's own admission that calls had risen⁹.

Who benefits. The NatWest case involved a customer who needed support managing his account and could not get past the chatbot⁶. It is a single case and says nothing about frequency.

Dependence on people. Commonwealth Bank relied on overtime and team leaders to answer calls⁹, and both ombudsman decisions turn on access to a person⁶⁷. Part of any gain therefore depends on staffing and process, not on the AI.

What is actually AI. Bank of America describes Erica as an "AI-powered" assistant without naming the method², and the CFPB reports that rule-based bots remain common¹. Only NatWest states generative AI use, in 21 journeys, without giving the share of its 12.9 million conversations that they handle⁴.

No withdrawal of a customer-facing banking assistant and no regulator enforcement action against one was found in the sources opened. The Commonwealth Bank reversal concerned a staffing decision, not removal of the voice bot⁸⁹.

Phase: In early use — why

The hypothesis was In wide use. Usage is wide, but only confirmed machine learning counts towards the phase. Bank of America does not state Erica's method², the CFPB reports that much of the industry uses rule-based bots¹, and NatWest confirms generative AI in 21 journeys with no stated share of traffic⁴. The scale of genuine AI use cannot be established, so the lower plausible phase is chosen. The phase would move up if operators published the machine-learning share of conversations and resolution rates.

What to watch next

• Publication of resolution or containment rates by Bank of America or NatWest; none appear in the sources opened²³⁴.
• Results or wider release of NatWest's Cora Spending Chat, in beta with selected customers in the retail banking app⁴.
• The first regulator or ombudsman action citing the AI disclosure duty reported to apply from 2 August 2026 against a bank assistant¹¹.
• Further Financial Ombudsman Service decisions on chatbot access, following the NatWest decision⁶.

Questions to ask before adopting

• Which parts of the assistant use machine learning, and what share of conversations do they handle?
• What share of conversations is resolved without a person, measured by whom and against what control?
• Can every customer reach a person in one step, including customers who need extra support?
• How will call volumes and staffing be measured independently after launch?
• Who is liable, and who answers to the regulator, when the assistant gives a wrong answer?