NatWest to trial voice AI spending insights with RBS customers

NatWest Group has announced plans to trial a conversational AI tool that turns a customer's own transaction data into spoken or written spending insights, complete with interactive graphs. The trial, disclosed on 25 September 2026, will run first through Royal Bank of Scotland (RBS), one of NatWest Group's retail banking brands.

The tool, called Spending Insights, is meant to let a customer ask about their spending by voice or text, interrupt, change subject or ask a follow-up question, and receive a personalised breakdown of habits, budgeting and possible savings. NatWest says the product was built by its Chief AI Research Office and runs on what it describes as a proprietary NatWest small language model.

For RBS customers, this matters because it is a planned test of whether generative AI can be trusted to interpret something as consequential as a person's own bank statement. NatWest has not yet said when the RBS trial starts, how many customers will take part, or how those customers will be chosen. That gap sits at the centre of the questions the trial raises about accuracy, privacy and support for people who rely on their bank to get this right.

How the RBS trial is meant to work

According to NatWest's announcement, the tool is designed to handle a natural back-and-forth, letting a customer interrupt, switch topic or return to an earlier point in the conversation. Output is meant to be interactive, presented as graphs covering spending patterns, budgeting and financial habits, with pointers to possible savings.

Beyond that, several practical questions remain open: whether customers must opt in, whether voice functionality can be switched off in favour of text only, and how privacy information will be presented to those invited to try it.

What NatWest means by its own small language model

A small language model is, in broad terms, a version of the technology behind familiar AI chatbots that is built or tuned to be smaller and more specialised than the largest general-purpose systems. NatWest attributes development of Spending Insights to its Chief AI Research Office and describes the underlying technology as a proprietary NatWest small language model.

Beyond that description, NatWest has not published the model's architecture, parameter count, training data, hosting arrangements, or evaluation results, nor has it said whether any third-party model or service supports parts of the system. The bank's own announcement is inconsistent even in its wording: the summary of the release refers to "Small Language Models" in the plural, while the body text uses the singular. It is not established from the public material whether Spending Insights runs on one model or several.

Not NatWest's first spending chatbot

Spending Insights is not NatWest's first attempt at letting customers talk to an AI system about their money. An earlier Cora pilot let selected customers ask natural-language questions about recent spending in the mobile app. NatWest disclosed in April 2026, in a separate announcement about home-buying guidance in ChatGPT, that this pilot involved 25,000 customers.

That 25,000-customer figure belongs to the earlier Cora spending pilot. It is not the disclosed size of the new RBS trial, which adds the announced voice-and-text conversational functions, and for which no cohort figure has been given. Readers should treat the two as related but distinct products.

The separate Cora fraud agent

Alongside Spending Insights, NatWest has launched a Fraud Triage Agent within Cora, aimed at suspicious credit-card transactions rather than everyday spending. It gathers a customer's account of what happened and directs them towards the appropriate fraud, scam or dispute route. NatWest's available material does not say whether the agent itself determines whether fraud occurred, or who is liable for any loss.

NatWest says the Fraud Triage Agent has handled nearly 19,000 customer conversations since launch. The bank has not stated when the agent launched or the measurement cut-off date, and the figure counts conversations rather than unique customers, resolved cases, or correct routing outcomes. It should not be read as evidence about Spending Insights, a separate product covering different interactions.

NatWest's general guidance on Cora says conversations are stored for monitoring and quality, and logged-in conversations carry the same security protection as those services. That guidance describes Cora broadly; NatWest has not published trial-specific detail on how Spending Insights will store voice recordings, transcripts, prompts or generated graphs.

Privacy questions raised by transaction-aware AI

A tool reading a customer's transaction history and answering open-ended questions touches directly on UK data-protection law. Under the data-minimisation principle, an organisation using AI must not process more personal data than necessary, even if extra data could improve the model. NatWest has not disclosed which transaction fields or conversation records Spending Insights will process, whether prompts or transcripts train any model, or whether data leaves NatWest-controlled infrastructure.

There are further unanswered questions about household and shared finances: how the tool would handle a joint account, and how it would treat transactions revealing sensitive matters such as health or relationships. None of this is addressed in NatWest's published material.

Separately, changes to UK data-protection law effective 5 February 2026 replaced the old UK GDPR Article 22 with Articles 22A to 22D. The safeguards relevant here — information about a decision, a chance to make representations, human intervention and a route to contest it — are set out in Articles 22A to 22C, and apply where a significant decision is based solely on automated processing. NatWest describes Spending Insights as an information and insight tool rather than one making credit, pricing or eligibility decisions, so these safeguards would not automatically apply unless that changed.

Accuracy and the risk of false confidence

Getting the facts right matters more in banking than in most everyday chatbot uses, since a wrong answer can affect how someone manages money they cannot afford to lose track of. The ICO draws a relevant distinction: the legal accuracy principle for personal data is not the same as a model's statistical accuracy. Generative AI outputs need not be 100% statistically accurate to comply with data-protection law, but personal data used must be accurate enough for the purpose, and inaccurate personal data must be rectified or erased where required.

FCA research published 30 May 2025, based on two pilots using GPT-3.5 and GPT-4 rather than any NatWest model, found large language models can simplify complex financial information, but said validating outputs needs human judgement alongside automated tools, and that effectiveness depends on the customer journey's design. That is sector context, not a finding about NatWest's own, untested system.

NatWest has not published error rates, transaction-categorisation accuracy, or the frequency and severity of mistakes in Spending Insights. Nor has it explained how a customer could correct a miscategorised transaction, challenge a wrong inference, or reach a human reviewer. A reader relying on this tool to manage a budget faces real financial risk if an insight is inaccurate and acted on without independent checking.

Consumer Duty and human support

AI in financial services is not a regulatory blank space in the UK. The FCA said, in guidance published 8 September 2025, that it does not plan a separate AI rulebook and instead expects firms to apply existing, outcomes-based frameworks — including the Consumer Duty and senior-management accountability rules.

The Consumer Duty applied to open products from 31 July 2023 and closed products from 31 July 2024. It requires firms to deliver good outcomes, avoid foreseeable harm, support customer understanding, and provide suitable support. The FCA has told retail banks using AI to communicate analytics-driven decisions understandably, monitor outcomes and intervene quickly when problems emerge. Whether these duties bite on a given feature of Spending Insights depends on the regulated activity involved — detail the announcement does not address.

Access to a person remains part of that picture. NatWest's own commissioned research — a survey of more than 2,400 NatWest customers, benchmarked against 1,800 UK consumers — found that surveyed NatWest customers ranked access to a real person as the most important AI trust factor (81%), ahead of confidence in 100% reliability (77%), clarity about data handling (77%) and independent regulatory oversight (75%). These are NatWest's own figures; fieldwork dates, question wording and sampling method were not set out.

Sector-wide, a joint Bank of England and FCA 2024 survey found regulated firms viewed data privacy, data quality and data security as the three largest current AI risks — a comment on the sector generally, not on NatWest's tool specifically.

NatWest says its use of AI generally is governed by an Artificial Intelligence & Data Ethics Code of Conduct. A governance commitment of that kind is not, on its own, evidence that a specific tool performs accurately or fairly — the announcement includes no trial-specific audit, model card or independent assessment of Spending Insights.

What NatWest has and has not disclosed

What to watch next

NatWest's announcement is, for now, the only supplied first-party account of the RBS trial; readers wanting the primary description should consult NatWest Group's newsroom announcement rather than commentary about it. NatWest describes Spending Insights as an information and insight tool but has not explained how it will distinguish information, generic guidance and a personal recommendation, particularly when suggesting possible savings. Readers should look for NatWest's trial-specific privacy material once published, and direct questions to NatWest or RBS directly. What remains to be published — a start date, cohort details, retention periods, accuracy testing and a route to correct or escalate a wrong answer — would let customers and outside observers assess the safeguards and outcomes behind the trial, though whether the Consumer Duty applies to any particular feature still depends on the regulated activity involved and NatWest's ability to determine or materially influence that outcome.