Lloyds Bank's AI cost-cutting plan has a hidden tab: who pays when the machine gets it wrong?

Key points

• Lloyds Bank announced a £2 billion cost-cutting plan on 30 July built around AI-driven automation.
• Dr Gleb Tsipursky argues the bank's productivity figures only count time saved for the worker using the AI tool.
• Hidden costs include colleagues checking invented facts, fixing customer messages and explaining rejected loan applications.
• A single team can appear more productive on paper while the actual workload shifts to other parts of the bank.
• No public breakdown of these downstream costs has been published.

When Lloyds Bank unveiled its plan to cut £2 billion in costs on 30 July, the numbers looked clean. What they didn't capture was harder to see.

Dr Gleb Tsipursky, a decision-making and AI risk researcher, argues the headline figures leave out a critical question: when the AI gets something wrong, whose time absorbs the cost?

What does that hidden cost actually look like?

A colleague reads back a customer letter and catches a made-up interest rate. A branch worker calls a customer to explain why their mortgage was rejected because of a system error. An escalation queue appears that wasn't there before.

AI tools that draft text or process applications can produce what researchers call hallucinations: confident-sounding answers that are factually wrong. Catching those errors takes human time, and that time belongs to somebody else in the organisation.

Tsipursky's argument is a straightforward accounting problem. Banks measure minutes saved by the employee directly using the AI. They don't measure minutes spent by everyone cleaning up after it. One team posts a productivity gain; the risk and effort move sideways.

Should bank customers be worried?

Not dramatically, but there are things worth watching. If Lloyds' AI drafts a letter about your account, check any figures or deadlines before acting on them. If an application is rejected and the reason feels vague, ask for a human explanation in writing.

The real concern isn't a financial crisis. It's that errors diffuse across many small interactions, each costly or frustrating for the individual customer, while the bank's internal metrics record only the savings. Robert Charette, whose decades of work translating software risk into plain English we reflected on in August, called this kind of invisible cost the part that always shows up late.

Lloyds hasn't published a breakdown of how it will track the downstream human cost of its AI programme. Until it does, the £2 billion figure tells only half the story.

What should readers watch for?

Four things deserve attention if you're a Lloyds customer or work at any bank rolling out AI tools: letters quoting specific rates or deadlines without a named human author; rejection notices with no clear appeal route; tasks that now seem to require two people where one managed before; and any process where the AI's output reaches you without a human review step between the machine and your money.

The accounting fix Tsipursky calls for isn't complicated. It's just inconvenient for banks whose savings projections don't budget for the cleanup.