Executive Summary

Key numbers

The first two give how many roles go and where they sit. The last two are what the company published about bringing AI into those same places, and what the restructuring costs.

Source: Barclays Q3 2026 results (October 22, 2026), Anthropic announcement (October 1, 2026)

The announcement came in a good quarter. Third-quarter profit before tax reached £2.3 billion, up 12% on the same period last year, and equities trading income jumped 47%. Fixed income, currencies and commodities fell 19%, which Venkatakrishnan singled out as a disappointment, but across the group this was not a quarter in which the money ran short. Barclays announced a share buyback of up to £500 million on the same day.

The location of the roles going was disclosed fairly clearly. Cross, the chief financial officer, said about 60% of the actions related to support functions, developers included. What the company said is that support functions account for 60%, not that developers account for 60%. Developers are one occupation inside that bundle, and how many of them there are is not in the announcement.

The sentence Cross used on the results call ran like this. "These actions involve an expected reduction of circa 5,000 full-time employees. Around 60% of the group actions relate to support functions, including fewer developers as AI makes our people more productive." The 60% attaches to actions, not to people. Reporting carried it over as 60% of 5,000, a little more than 3,000 roles, and this article follows the same rough arithmetic, but the company never pinned the 60% to a headcount.

Gathering the items that carry a figure in the announcement and the results material gives this.

Reading the table downward, one thing stands out. How many roles go, what that costs, and how much the bank means to save are all written as figures. There is no row for how many of those roles went because AI took the work. Filings do not ask for such a row, so its absence is unremarkable in itself.

Causation was therefore explained in words rather than figures, and the weight shifted with the speaker. Cross said the changes in the third and fourth quarters related both to the simplification of the organization and to some of the impacts the bank was starting to see from AI. Venkatakrishnan was more guarded. He said Barclays would employ AI at scale and then added that "AI is the icing on the cake, not the cake," since the infrastructure has to be built before the benefits arrive. What had driven the efficiency gains so far, he said, was platform modernization, system consolidation and simpler processes.

The icing image had evidence behind it. The improvements the two of them cited on the same call did carry before-and-after figures. On the mortgage broker platform, application times fell from 45 minutes to under 15 minutes, and broker net promoter score moved from minus 51 to plus 70. Customer-facing technology cut call center volumes by 10% over the past year, and a change to the branch staffing model brought employees per branch down by around 10% over the same period.

The first thing those four figures share is that they compare a before with an after. The second is attribution. The mortgage gain was put down to technology investment, the call center to customer-facing technology, the branches to a change in the operating model. Not one of the four was presented as an achievement of AI. That is the context for the chief executive's remark about building the infrastructure first. It also means the AI column is not blank because this is an organization that fails to keep records.

A month before the cuts, on September 22, 2026, Barclays Research published a report analyzing the effect of AI on labor. According to the reporting, it is called the AI Automation and Amplification Framework, or 3A Framework. Its distinguishing move is to take skills rather than whole occupations as the unit of analysis, mapping 104 cross-occupational skills from the US Department of Labor database onto some 830 occupations and scoring exposure to cognitive AI and physical AI separately.

The message the report led with ran the other way, toward reassurance. Only about 20% of core work skills are ones AI can substitute for at a high level, it said, and the story of wholesale replacement is overdrawn. What spreads far more widely, in its reading, is amplification: AI assisting human work and raising productivity.

The same report also drew the distribution of that exposure. Occupations with high cognitive AI exposure cluster in computing, engineering and science, and those with high physical AI exposure cluster in construction, manufacturing, installation and maintenance. A month later, the 60% share of the actions the company said it would take sat in the first of those clusters, the support functions that include developers.

Barclays has never said this report fed into the decision on headcount, and it has offered no account linking the two. This article does not claim the report directed the cuts either. The report itself is not open to the public, so what is set out above rests on reporting. Even so, it is worth recording that one company issued an analysis saying replacement is overdrawn and, a month later, an announcement cutting roles in the category that analysis had marked as highly exposed.

Both documents can be true at once. Even if the substitutable skills are 20% of the whole, a concentration of that 20% in particular occupations will shrink those occupations. The difficulty is that neither document shows "AI did this work instead" as a measured value. One is an estimate of exposure, the other a plan for headcount. No observational record bridging them has ever been published.

The place where Barclays set out AI figures in the most detail was not the cuts announcement but three weeks before it. On October 1, 2026, Anthropic announced an expansion of its work with the bank and wrote the internal deployment and the plans out as numbers. Three values are public.

All three measure the size of the usage. How many people use it, how many items passed through, what percentage it will reach next. What is nowhere in the announcement is the comparison with how the same work went when people did it: how much the time to handle one item fell, how errors and rework shifted, how a team using it compares with one that does not. The same goes for a contract value and a saving.

The rule separating the left box from the right is simple. The left falls out of an adopting organization tallying its own activity. The right requires the same work to have been observed under two conditions. The second cannot be reconstructed after the fact. Ask "so how much faster is it" once the rollout is finished and there is no way to answer if no record of the earlier state survives.

The same announcement contains a passage where the bank writes down, in its own words, that effect is what should be measured. Anne Marie Darling, group co-chief operating officer at Barclays, said: "The true measure of any technology is the impact it has on customers, clients, and colleagues, and the outcomes it enables." An announcement that defines the measure that way carries no value measured by it.

The other group co-chief operating officer, Craig Bright, said: "Software engineering and cyber security are both being reshaped by increasingly capable AI systems," describing the use of Claude for modernizing legacy platforms and improving software quality. That was three weeks before the cuts. Between the occupations the September report marked as highly exposed and the 60% of October's cuts, there is one more occasion on which an executive at the company named those occupations out loud. Here too, what Bright named was the direction of the reshaping, not its scale.

There is no basis for concluding that Barclays has been lax about measuring effect. There are many reasons to hold material internally even when it exists. Within what can be checked from outside, though, every figure this company has attached to AI is an adoption figure. The comparison figures from section 1 were attached to the mortgage platform, the call center and the branches. The bank has also said its executive committee will invest about £150 million in skills by 2030 and will prioritize filling advertised openings with existing employees. Where the money goes is written down. The column measuring what the money changed is still empty.

Headcount planning at a bank is not what Pebblous works on. The gap this announcement exposes, though, has the same shape as the one we meet daily in the work of data quality. When a team that brought AI in writes its report six months later, what it usually holds are user counts and call counts. A record of how the same work was handled before the tool arrived almost never exists.

Adoption rates show how fast a tool spread through an organization. On their own they say nothing about how much work was displaced. Building the material that joins the two means leaving three things behind before the rollout starts.

• A baseline for the same work before adoption. Time to handle one item, share returned as errors, number of times a person had to rework it, all measured before the tool arrives, or the comparison never holds
• A record of the same period for teams that did not adopt. Looking only at firm-wide metrics makes it easy to read the effect of the economy or a reorganization as a win for AI
• A map from roles removed to tasks automated. Which tasks in which jobs moved to which tool has to be on paper for the sentence "AI replaced it" to have anything underneath it

The point is that none of the three can be built after adoption. A baseline is a past state, a control group requires a team that held off, and a task map is accurate only at the moment of the switch. Without them the effect of AI comes down to who speaks with more conviction. That is probably why the chief financial officer and the chief executive at Barclays gave it different weight on the same day.

Venkatakrishnan touched the same point in a remark he made to Bloomberg back in May, when he said that what people do are jobs and that a job is much more than a collection of tasks. Automation happens at the level of tasks and headcount is counted at the level of jobs. Something has to translate between the two units, and the third record above is what does the translating. Without it, the jobs that went and the tasks that were automated never meet in the same table.

Thanks for reading this far. The figures in this article were checked against the full Barclays third-quarter results call of October 22, 2026 and the material Anthropic published on October 1, and the September report section rests on reporting because the original is not open to us. If you report on an AI rollout inside your own organization, we would be glad to hear which number you write down next to the adoption rate.

Pebblous Data Communication Team
October 24, 2026

Sources cited in this article, grouped into primary sources and industry/news coverage.

R.1Primary Sources

• 1.Barclays. (2026). "Q3 2026 Earnings Call Transcript." The Motley Fool
• 2.Anthropic. (2026). "Barclays scales Claude to upgrade operations and improve client experience." anthropic.com

R.2Industry & News Coverage

• 3.CNBC. (2026-10-22). "Barclays to cut 5,000 jobs as AI starts creating efficiency." cnbc.com
• 4.Bloomberg. (2026-10-22). "Barclays to Cut 5,000 Jobs as AI Starts Creating Efficiency." bloomberg.com
• 5.Techstrong.ai. (2026-10-22). "Barclays to Cut 5,000 Jobs in AI-Driven Cost-Reduction Push." techstrong.ai
• 6.Banking Dive. (2026-10-22). "Barclays to cut 5,000 employees this year." bankingdive.com
• 7.Tech Times. (2026-10-23). "Barclays Cuts 5,000 Jobs After Its Own AI Research Flagged Developer Roles." techtimes.com
• 8.AI Times. (2026-10-23). "Barclays cuts 5,000 jobs as AI adoption grows, concentrated in developer and support functions." (Korean) aitimes.com