Basil C. Puglisi

Artificial Intelligence (AI) & Digital Marketing Since 2009

Barclays and Amex Cut 5% Each for AI. Neither Said Who Will Check the Work #AIg

by Leave a Comment

Barclays Q3 2026 results slide titled Structurally reducing the cost base to generate capacity for investment, listing about 5,000 FTE reductions affecting about 5% of employees, about 60% in support functions including developers, and organisational simplification and technology change efficiency driven by AI
Image: Slide 10, “Structurally reducing the cost base to generate capacity for investment,” from the Barclays Q3 2026 Results Presentation, October 22, 2026. Source: Barclays PLC.

Two financial companies told investors last week that about one job in twenty is going away, and both pointed at AI. Neither one said who will check the work the remaining people and their AI tools produce once those jobs are gone.

Barclays went first. Its third-quarter results presentation on October 22 lists an expected reduction of about 5,000 full-time employees, roughly 5% of the bank, with about 60% of the cuts in support functions, including developers. The slide names the driver in one line: “Organisational simplification and technology change efficiency driven by AI.” On the results call, the bank’s finance director said the cuts include “fewer developers as AI makes our people more productive.”

American Express followed the next day. Its quarterly filing records a $336 million restructuring charge, paid for with part of the gain from selling its stake in Global Business Travel Group. Chief executive Stephen Squeri told analysts on the earnings call that it eliminates about 5% of positions globally, and chief financial officer Christophe Le Caillec told CNBC the point was to make the company “leaner, more agile, and also to prepare for AI.”

Both companies said something right

Credit goes where it’s earned. Barclays chief executive C.S. Venkatakrishnan closed his remarks with a line every board should borrow: “AI is the icing on the cake, not the cake.” The bank’s own numbers back a measured story. It credits customer-facing technology with cutting call centre volumes by 10% in a year, and changes to branch staffing with about 10% fewer employees per branch.

Le Caillec got the sequence right too. “You don’t want to automate processes that are not optimally structured,” he said. Redesign the process first, then automate it. That’s the right order, and plenty of companies skip it. Squeri added a boundary for outside AI: Amex is “not going to let assistants run rogue through our system,” and will admit agents only through approved API or MCP channels.

Wider spans, fewer developers, same volume of output

Here’s the part the announcements leave out. Le Caillec described the method as cutting layers and increasing spans of control “so that we can automate and take advantage of AI in the coming quarters and the coming years.” A wider span means each manager oversees more people. Add AI and each manager also oversees more machine output: drafted letters, reconciled accounts, generated code. The review load per person rises in the same quarter the reviewers get fewer.

Barclays’ developer line raises the same question from the other side. If AI makes the remaining developers more productive, more code ships per person. Automated tests and scanners can confirm that code runs and matches known patterns. That’s Responsible AI: software verifying software, with no person reading the result. They can’t decide whether a change to a payments system is the right change. That decision belongs to AI Governance, which needs a named person who can approve the change, send it back for edits, or stop it, and who answers for whichever call they make.

Amex already knows regulators expect that person. Its own 10-Q warns that new AI laws are imposing “obligations related to data use, recordkeeping, transparency and human oversight.” The company guards its front door against rogue agents. In the same quarter it is removing management layers, and in most companies those layers are where the human review of everyday output actually happens.

The payroll data shows where the cuts land

The day before Barclays reported, the Canaries Dashboard from ADP Research and the Stanford Digital Economy Lab showed employment in the occupations most exposed to AI down 0.5% from a year earlier in September, while the least exposed grew 1.0%. For workers aged 22 to 25 in highly exposed jobs, employment fell 3.7%, the 36th straight month of year-over-year decline. Workers aged 26 to 30 in those jobs were down 2.8%.

Basil has written about what that dashboard can and can’t tell us. In The On-Ramp Problem, he argued that the dashboard is a population-scale diagnosis that reads the labor outcome after it lands, while the variable that decides it, whether AI augments a worker or replaces the task, has to be measured one person at a time. A 5% cut announced on an earnings call has the same blind spot. It reports the jobs removed. It doesn’t report the judgment removed with them.

Four questions before an AI-justified cut

Basil’s Checkpoint-Based Governance, which he applied to the future of work in Human-AI Collaboration Needs More Than a Human in the Loop, differs from ordinary human-in-the-loop review by requiring binding authority, a decision record, and a measured check on rubber-stamping. Applied to a workforce reduction made in AI’s name, it asks four things before the notices go out:

Which tasks is the AI taking, and what evidence shows it does them at today’s quality? A productivity estimate is a forecast, not evidence. A pilot with error rates measured against the human baseline is evidence.

Who reviews that output after the cut, and how many hours does the review take? If the answer is the same managers with wider spans, put the hours in the plan and staff them before the reduction, not after the first incident.

What reverses the decision? Set a written trigger, such as rework or complaint rates rising past a stated level in the affected process, that reopens the staffing call. A cut with no reversal condition was never really a judgment about AI.

Did the people doing the work get asked? Barclays’ slide notes its reductions are subject to collective and individual consultation as required. Used well, that consultation is the best source of the tasks the AI can’t handle yet, because the people doing them know exactly where the exceptions live.

The number worth adding to every quarterly update is review hours per unit of AI output, before and after the reduction. If that figure falls while output climbs, the savings are partly coming from checking less.

Barclays expects more than £1.1 billion in gross efficiency savings in 2027, and Amex expects a leaner company in the coming quarters. Both numbers will be easy to report. The harder report is the one neither company has promised: who now reads what the AI produces, and how often they say no.

Sources

FAQ

How many jobs is Barclays cutting, and how does AI figure in?

Barclays’ Q3 2026 results presentation on October 22, 2026 lists an expected reduction of about 5,000 full-time employees, roughly 5% of employees, with about 60% in support functions including developers. The bank attributes the cuts to organisational simplification and technology change efficiency driven by AI, and its finance director said the actions include fewer developers as AI makes staff more productive.

Why is American Express cutting about 5% of its positions?

American Express took a $336 million restructuring charge in the third quarter of 2026, funded with part of its gain from selling its stake in Global Business Travel Group. CEO Stephen Squeri told analysts it eliminates about 5% of positions globally, and CFO Christophe Le Caillec told CNBC the goal was to make the company leaner, more agile, and prepared for AI by reducing layers and increasing spans of control before automating.

What does the Canaries Dashboard show about AI and jobs in September 2026?

The Canaries Dashboard from ADP Research and the Stanford Digital Economy Lab reported on October 21, 2026 that employment in the most AI-exposed occupations fell 0.5% year over year in September while the least-exposed grew 1.0%. Workers aged 22 to 25 in highly exposed jobs saw a 3.7% decline, the 36th consecutive month of year-over-year contraction.

What is the governance risk in widening spans of control while adopting AI?

Wider spans mean each manager oversees more people, and AI adds more machine output to review. Automated checks can confirm that output runs or matches patterns, but deciding whether it is right needs a named person with authority to accept, modify, or reject it. Cutting layers can quietly cut that review capacity while output volume rises.

What should a company answer before cutting jobs in AI’s name?

Under Checkpoint-Based Governance it should answer four questions: which tasks the AI takes and what measured evidence shows it performs them at current quality, who reviews the output afterward and how many hours that takes, what written trigger would reverse the decision, and whether the affected workers were consulted about the tasks AI cannot yet handle.

What metric shows whether AI-driven savings come from checking less?

Track review hours per unit of AI output before and after the reduction. If that figure falls while output climbs, part of the reported savings may come from reduced human review rather than from AI doing the work well.

AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated

Subscribe to Blog via Email

Enter your email address to subscribe

Join 9,585 other subscribers

Reader Interactions

Leave a Reply