What a 7.4% chatbot resolution rate means for UK banks

A vendor-produced study says just 7.4% of chatbot interactions in a combined "financial services and insurance" test category achieved the goal the test set out to check, against a chatbot adoption rate of 64.2% in the same category. The findings, published by the customer-experience software company Parloa on 22 April 2026, have circulated as evidence that bank chatbots fail most customers. That framing goes further than the underlying research supports. Parloa's category combines financial services and insurance, a broader grouping whose company composition Parloa has not disclosed, its sample was drawn from global enterprises rather than UK institutions, and no UK-only figure has been published. For UK banks, the more solid story is not the precise percentage, but what the Financial Conduct Authority's (FCA) Consumer Duty already requires of automated customer support, regardless of how any firm scores in a vendor's benchmark.

UK retail banks, building societies and other FCA-regulated firms that route customer enquiries through chatbots are affected, along with the customers who rely on those channels. It matters now because the Consumer Duty's consumer-support rules already apply to every support channel a firm uses, including one it has outsourced to a third-party AI vendor, and because the Bank of England and FCA's most recent survey found AI already in widespread use across UK financial services.

What Parloa actually tested

Parloa says its 2026 report, The State of Agentic CX, used AI agents, with human involvement, to examine more than 10,000 enterprise websites, carry out nearly 4,000 chat interactions across more than 800 companies in 27 industries, and test 100 telephone trees. The industries covered include banking, insurance, travel, retail and healthcare, drawn from what Parloa describes as a global enterprise sample. The publicly available methodology does not disclose a UK-only subset, does not state how many of the roughly 4,000 chat interactions fell into the financial-services-and-insurance category, and does not list which banks or insurers were tested. Those are the details a reader would need before treating 7.4% as a description of UK banking specifically, rather than of Parloa's global test category.

The adoption-resolution gap

Within that combined financial-services-and-insurance category, Parloa reports three figures side by side: 64.2% chatbot adoption, 7.4% of tested interactions achieving the stated goal, and 65.7% of the chatbots it could classify as rule-based systems. Separately, across the full 27-industry sample, Parloa reports that 8.9% of chatbot interactions achieved the stated goal.

The 7.4% and 8.9% figures are not two competing estimates of the same thing. They cover different populations, one sector combination against the whole 27-industry sample, so the gap between them is not evidence of finance-specific underperformance on its own. Parloa's other material puts rule-based prevalence differently again across its full sample, a reminder that the underlying populations shift between statistics and that comparisons across Parloa's own tables need care.

The broken hand-off question

Parloa's figures continue into what happens when a chatbot cannot resolve a query itself. Across the full study, Parloa reports that only 10.1% of attempts to escalate a chat to a human reached one, implying that 89.9% of escalation attempts did not. The Fintech Times, reporting on 26 September 2026 on an interview with Parloa's chief executive, Malte Kosub, presented the 89.9% failure figure as a finding about financial services specifically. Parloa's own published key-findings material presents the 10.1% success rate as a result across the whole study, not as a financial-services-only statistic. Which reading is correct cannot be settled from the material available. It would need the underlying sector-level table or dataset, which has not been published.

Why the numbers need caution

Several features of the research call for caution before any UK bank cites it, or before any UK customer treats it as a description of their own bank. Parloa sells agentic AI customer-service software, so a report showing that other approaches underperform makes a case for its own product. That commercial interest does not make the research wrong, but it does mean the figures need independent context rather than being repeated at face value. No public breakdown discloses the number of financial-services-and-insurance companies or interactions behind the 7.4%, 64.2% and 65.7% figures, nor whether the test tasks were realistic for an unauthenticated website visitor. No confidence intervals, error margins or statistical-significance testing accompany the industry-level figures, and no independent replication of the 7.4% result has been published.

What Consumer Duty requires of UK banks

For UK-regulated firms, the more durable analysis sits with the FCA's Consumer Duty rather than with any single vendor benchmark. The Duty's consumer-support rules took effect for open products and services on 31 July 2023 and were extended to closed products and services on 31 July 2024. Under PRIN 2A.6.1R, those rules apply regardless of the channel a firm uses to support customers, and they apply to support a firm has outsourced to a third party. Routing enquiries through an externally supplied chatbot does not, by itself, move the regulatory responsibility to the vendor.

PRIN 2A.6.2R requires that support meets retail customers' needs, including customers with characteristics of vulnerability, allows customers to use products and services as they could reasonably expect, and avoids creating unreasonable barriers at any point in a product's lifecycle. FCA guidance describes unreasonable barriers as steps that are unreasonably onerous, time-consuming, complex or difficult to understand, and that can cause delay, distress or inconvenience. That description could cover a chatbot that repeatedly fails to escalate a customer to a person, or that obstructs a complaint, a switch or a cancellation, though whether it does depends on the individual customer's actual outcome, not on a hypothetical.

The FCA has said it does not plan to introduce a separate regulatory regime for AI in financial services, relying instead on existing frameworks: the Consumer Duty, the Senior Managers and Certification Regime, and firms' governance and control requirements. The reviewed FCA material does not prescribe a minimum chatbot resolution rate or require a human agent to be available for every interaction. Whether an automated route counts as an unreasonable barrier is assessed case by case against the facts of the customer's outcome, rather than against a fixed threshold.

From containment to outcome evidence

Parloa's reported chatbot adoption rate, 64.2% for the combined category, is a different measure from the customer outcome the Consumer Duty is concerned with, and it should not be read as evidence of how many customer issues automated support actually resolved. The FCA's good-practice material on the consumer-support outcome, and its July 2026 article on outcomes monitoring, both describe an expectation that firms understand customers' actual experience, identify where monitoring shows poor outcomes or a risk of harm, and act on it in good time. For a chatbot programme, evidence that would speak to that expectation could include goal completion, repeat contact, abandonment, successful escalation, waiting time, complaints, and how each of these varies across customer groups, including customers with characteristics of vulnerability. The FCA does not prescribe this exact set of chatbot metrics; it is a reasonable reading of what understanding an actual outcome would require in this context, not a quoted rule.

The reason this matters is not confined to a single vendor's test. The Bank of England and FCA's November 2024 survey found that 75% of responding UK financial firms already used AI, with a further 10% planning to within three years, and that around one-third of reported AI use cases were implemented by third parties. Whatever share of a firm's overall AI use that represents, Consumer Duty obligations continue to apply when customer interactions are outsourced to a third party.

Could agentic AI improve the position?

Parloa's report argues, and its commercial pages make the same case, that more capable agentic systems, able to complete a task or hand a customer to a person with context intact rather than simply answer with a script, could close some of the gap the study describes. That argument is unverified against the research itself. The benchmark tested existing chatbots as they perform today; it did not independently test whether replacing them with agentic systems produces sustained, safe improvements in a live, regulated banking environment. Open questions include how an agentic system that can authenticate a customer or execute a transaction would be governed for fraud risk, incorrect or fabricated information, privacy, audit trails and erroneous transactions, and whether any comparative UK trial has tested these systems against the outcomes the Consumer Duty is concerned with. Answering them would need independent evidence, not a vendor's own benchmark of the technology it sells.

What to watch next

The FCA's Handbook sets out the consumer-support rules in PRIN 2A, and the regulator's published material on its approach to AI and on outcomes monitoring remain the primary references for how existing rules apply to automated support. What is missing from the public record, and would change how far this story can be told, is a UK-specific breakdown of chatbot performance, sector-level detail from Parloa's own dataset, and independent testing of whether newer agentic systems perform differently from the rule-based systems the study describes.