What AI at scale actually looks like: 16,000 colleagues and 120,000 emails a day

Barclays has 16,000 colleagues using one AI system and about 120,000 client emails a day passing through another. Neither of them decides anything.

The first finds an answer that already exists somewhere in the bank’s own documents. The second reads an incoming request and works out which desk should handle it. That is what “AI at scale” looks like inside a universal bank at the end of 2026, and we think it is the most useful fact an operator can carry into next year’s budget.

The forecasts say something else. Gartner expects agentic AI to resolve 80% of common customer service issues without human intervention by 2029, cutting operational costs by 30%. Set that against what a bank serving more than 20 million UK retail customers has actually switched on, and the distance is what interests us.

What is actually live

Barclays’ Colleague Knowledge Assistant has been running since 2025. It is a retrieval system. Staff ask it a question, it reads the bank’s own material, and it hands back the passage. More than 16,000 colleagues have adopted it, and it has handled over a million searches.

In Global Markets, Claude models classify, enrich and route incoming client email. Around 120,000 messages a day go through it, so that a request reaches the right desk carrying what an operations colleague needs to act on. Claude Code is expected to reach 50% of Barclays’ developers by the end of 2026, and a majority of its engineers in 2027.

Those are the vendor’s figures, published by Anthropic on 1 October 2026, and the release carries no cost, no error rate and no measured hours saved, so we read them as the shape of the deployment rather than its return. Before calling it a result we would want a cost per search. Anthropic’s commercial chief, Paul Smith, puts it as Claude helping colleagues “find answers for customers” and sorting email. That is a precise description of two pieces of plumbing.

Is Barclays being cautious?

It is being ordinary. The Bank of England and the FCA surveyed 118 regulated firms and found that 2% of AI use cases have fully autonomous decision-making. A further 2% run automation with dynamic models. Some automated decision-making appears in 55% of use cases, and firms rate 62% of their own use cases low materiality against 16% high.

The share was lower than we expected, and nothing in the deployment news of the following two years makes it look like a rounding error.

The survey was published in November 2024 and is still the only census of its kind in UK financial services, so the share has very likely risen; our guess is mid-single digits today, and we would rather say so than pretend the census is current. The direction of the finding is harder to argue with: three quarters of the firms were already using AI, up from 58% two years earlier, and almost none of it was deciding.

What they were using it for is just as plain. Optimisation of internal processes was named by 41% of respondents, cybersecurity by 37%, fraud detection by 33%.

Why does routing pay first?

Because sorting is the part of a workflow where the machine’s weakness does not matter. A misrouted email is noticed by the next person who reads it. A wrong retrieval is checked by the colleague who asked. Both failures are visible, cheap and recoverable in minutes, which is not something we would say about a credit decision or a suitability assessment.

The test we use is deliberately dull. Four properties, together, decide whether a workflow is a sensible first target, and Barclays’ two live systems have all four.

Volume matters most, because retrieval and routing save seconds and the return comes from the count. An inbox at 120,000 messages a day is worth automating badly. An inbox at 400 is not worth automating well.

What the regulators are now asking

The useful shift in 2026 is not about autonomy at all. At the Bank of England and FCA’s AI Consortium, co-chaired by Sarah Breeden of the Bank and David Geale of the FCA, members observed in June that human-in-the-loop “is widely relied upon” but “may not be practical in all contexts”. They proposed keeping agent actions inside clearly defined limits instead. And they moved the supervisory question: not whether one firm can deploy an agent safely, but whether the wider system can observe, test and contain its behaviour.

That is a harder question than most AI roadmaps answer, and we read it as the real change of the year. It also explains why the loop is still in place where the vendors are pushing hardest. Anthropic shipped ten agent templates for finance work in May 2026, published as plugins and cookbooks: KYC screening, month-end close, pitchbooks. Of all of them — including the ones that run unattended as Managed Agents — it says that users “stay firmly in the loop — reviewing, iterating on, and approving” before anything is filed or acted on. Its Opus 4.7 model leads Vals AI’s Finance Agent benchmark at 64.37%: state of the art, and a long way from a score you would sign an audit opinion against.

The ambition in the quotes is real, though. FIS’s chief executive Stephanie Ferris describes an agent that “compresses AML investigations from days to minutes”. BNY’s Leigh-Ann Russell talks of “new digital employees who work the case end to end”. Those are the same firms that keep a reviewer in the chain today.

On the forecast, we think the 80% will arrive in the channels where a mistake costs a refund, and that it will not arrive by 2029 where a mistake costs a complaint to a regulator. The split will not run along industry lines. It will run along error cost, inside the same firm, often inside the same team.

Which makes the operator’s question narrower than “what is our AI strategy”, and far easier to answer. Where in your business does a large number of requests arrive, need sorting rather than judging, and get acted on by somebody who would notice a mistake? Barclays found two, and we would expect most businesses of any size to have one. It is rarely the one on the roadmap.