TL;DR: As banks wire AI assistants into customer operations, software testing teams face a failure mode with no obvious owner: a single transcription error entering a summary that later feeds analytics, complaint reviews and model training. Benjamin Walker, chief executive of Colorado transcription firm Ditto Transcripts, calls it error inheritance.

The immediate prompt is EricaAssist, which Bank of America has just extended to over 18,000 customer service representatives. It summarises why someone is calling, pulls relevant records and proposes next steps inside three seconds, and the bank reports average call length down by close to a minute. Those numbers are real. The quality question sits underneath them.

Walker’s illustration is precise: a system that mishears a disputed sum, drops the word “not”, or attributes a remark to the wrong speaker produces a summary that reads polished and authoritative. “The danger is that nobody reopens the recording because the summary reads so confidently,” he said. “Three seconds is impressive, but banking decisions can follow a customer for years.”

The propagation path is what makes this different from an ordinary defect. Transcripts feed customer feedback analysis, staff training, complaint investigations and strategic planning. An uncorrected recognition failure repeated across records stops looking like a defect and starts looking like a customer trend — at which point it shapes decisions no one traces back to the microphone.

Industry data suggests the discomfort is widespread. Cambridge Judge Business School’s Global AI in Financial Services Report this year recorded unreliable outputs and hallucinations troubling 70% of institutions surveyed, with 40% naming data availability and quality as a barrier. ACA Group, polling over 200 financial services firms across the US, put adoption at 84% somewhere in the business, though active deployment averaged under 20% in compliance and roughly 5% in operations, held back by error rates, regulatory scrutiny and the absence of clean audit trails.

For UK firms the practical shift is from testing a model to testing a workflow. Regression testing, traceability and output validation matter every time a bank changes its speech recognition, its language model, its prompts or its knowledge base — because a subtle defect introduced at any of those points travels. Walker’s proportionality rule is the sensible starting position. “A routine service call, a fraud interview and a legal deposition do not carry the same consequences,” he said — so how much a human checks should track what the record will eventually be used for.

Looking forward: Expect this to surface in Consumer Duty and complaint-handling reviews before it surfaces anywhere else, since that is where an inherited error first meets a regulator with the power to ask how it got there.