Barclays has split its AI spend by job. Microsoft 365 Copilot is the general assistant on 100,000 desks, and Anthropic's Claude handles the narrow work that can be counted: a knowledge assistant, triage of about 120,000 client emails a day, and Claude Code for half the bank's developers by the end of 2026. If you run a large Microsoft estate, this is the clearest public template yet for buying a second model vendor. It also shows what to demand before you copy it, because Barclays has published how much the tools are used and nothing on what they produce.
Anthropic announced the expanded deal on October 1, 2026. Barclays expects Claude Code to reach 50% of its developers by the end of this year and a majority of its software engineers in 2027. Sixteen months earlier, in June 2025, the same bank rolled out Microsoft 365 Copilot to 100,000 colleagues after a 15,000-person pilot. Craig Bright, the executive quoted in both announcements, is now Group Co-Chief Operating Officer. The same executive fronted both deals, one vendor for each job.
What Did Barclays Give Each Vendor?
Barclays gave Microsoft the desktop and gave Anthropic three workloads with a clear input and output. The split looks like this, using only what the two announcements state:
| Microsoft 365 Copilot | Claude | |
|---|---|---|
| Users | 100,000 colleagues, global | 16,000+ UK colleagues (knowledge assistant); 50% of developers by end-2026 (Claude Code) |
| Job | General assistant inside Microsoft 365, plus a "Colleague AI Agent" for travel, policy and HR questions | Retrieval over internal knowledge for customer-facing staff; classifying and routing Global Markets client email; writing and modernising code |
| Published usage | Seat count | 1M+ knowledge searches since 2025; ~120,000 emails a day |
| Published outcome | None | None |
Microsoft's own write-up also describes "Colleague Front Door," an agentic dashboard accessed through Microsoft Viva for booking desks and annual leave. Darren Hardman, Microsoft UK's chief executive, framed Copilot as "the UI for Barclays AI." Microsoft cast Copilot as the front end for the bank's AI, and Barclays still sent the coding work and the email pipeline elsewhere.
On the Claude side, The Asian Banker's report describes the Global Markets job as classifying, enriching and routing about 120,000 incoming emails a day. The knowledge assistant is a retrieval-augmented generation (RAG) system, which means the model answers from documents it fetches out of the bank's own knowledge base instead of from its training data. Anthropic says it serves staff who look after more than 20 million UK retail customers.
Why Would a Microsoft Bank Buy a Second Model Vendor?
The answer is that a seat licence and a workload are priced and measured differently, and Microsoft's own licensing now says so. Microsoft now splits Copilot into two bills. Its [licensing explainer], dated September 25,(https://learn.microsoft.com/en-gb/microsoft-365/copilot/user-subscription-license-usage-based-billing) says the user subscription licence covers "everyday AI" such as summarising meetings and drafting documents, while Cowork, Code, Autopilot and frontier models run on usage-based billing that consumes Copilot Credits. We covered that change in detail in Microsoft's New Copilot Bills Autopilot Outside the $30 Seat.
So the coding and agent work your executives assume comes with Copilot is metered either way. Once both options are metered, the comparison is workload against workload, and a bank is free to pick a different vendor for each one.
The same Microsoft document says model selection inside the Copilot seat includes Anthropic's Sonnet "with fair use" and Opus "with limits." A Microsoft shop can already get Claude models through Copilot. What Barclays bought directly from Anthropic was a set of surfaces: Claude Code in the developer's terminal and Claude wired into back-office pipelines. If you are weighing the same choice, frame it that way. You are choosing where the model runs and who instruments it. Which lab trained it is a smaller question.
Bright's quote in PYMNTS' coverage names the target: using Claude "to help modernize legacy platforms, improve software quality and allow our technical experts to focus on the most complex challenges." Legacy modernisation is a job with a backlog and an end state. It is easier to fund as a workload than as a seat.
What Barclays Did Not Publish
Barclays has published adoption numbers and no results. There is no accuracy rate for the email router, no misroute or escalation rate, no answer quality score for the knowledge assistant, no cost, and no productivity figure for Claude Code. The bank's targets are stated as a share of developers using the tool, which measures reach. Whether the code got better or shipped faster is a separate question, and nobody has answered it in public.
The Wealth Advisor put the gap in its headline: "The Real Test Is Whether Anyone Notices." It also points out that Barclays' own economists have doubts. In July, Barclays Private Bank titled its AI Mid-Year Outlook 2026 "The productivity paradox" and wrote that AI's gains depend on "adoption, cost and the share of tasks it can address." A summary of Barclays analysts' July research reports they called evidence that AI makes workers more productive "unconvincing," and cites a figure of only 14% of US workers using AI daily for work in Q2 2026.
The deployment may well pay off. By the standard the bank's own research desk applies to everyone else, adoption figures alone do not show it yet.
The tooling to get that data exists, with caveats a bank should know. Anthropic's Claude Code analytics documentation offers lines of code accepted, suggestion accept rate, and, with a GitHub integration, merged pull requests containing Claude-written code. The same page says accepted lines exclude rejected suggestions and do "not track subsequent deletions." It also says contribution metrics are "not available for organizations with Zero Data Retention enabled," which a bank should check before relying on the dashboard. Ask for defect rate and cycle time as well; the dashboard leads with lines accepted.
What Does This Split Cost at List Price?
At list price the Copilot side is the larger, fixed line, and the Claude side is a small seat fee plus usage nobody has disclosed. Microsoft lists Microsoft 365 Copilot at $30 per user per month, paid yearly. Across 100,000 seats that is $36 million a year before any discount. Barclays' actual price is not public, so treat that figure as a ceiling.
Anthropic lists Claude Enterprise at $20 per seat per month plus usage billed at API rates, with Claude Code included. For every 1,000 developers, the seat fee is $240,000 a year. The usage line has no ceiling on the price list, and it is the number that decides whether Claude Code is cheap or expensive. Barclays has not said what it pays, how many developers it employs, or whether it caps per-developer spend. If you want to see how a usage line gets out of hand, our earlier piece on Claude Enterprise spend controls lists the settings to configure before rollout.
The arithmetic that matters for your own plan is simple. A seat licence is a cost per person whether they use it or not. A workload is a cost per unit of work. The email router is the cleanest example: 120,000 emails a day gives you a denominator, so cost per email can be compared directly with the cost of a person reading and routing it. Copilot on 100,000 desks has no such denominator, which is why M&T's Copilot disclosures left readers guessing how many staff actually used it each week.
What Does Human Oversight Mean for a UK Bank?
For a UK bank, "human oversight" is a supervisory expectation with a document behind it. Anthropic says Barclays runs Claude with "robust governance, security controls, and human oversight," but neither company says what that means for the email router: who reviews a misrouted client instruction, at what sample rate, and how fast.
The Prudential Regulation Authority's SS1/23 model risk management principles set the bar. The current version took effect on 23 April 2026 and applies to UK banks with internal model approval for regulatory capital. It covers "all types of models...regardless of technology" and explicitly includes managing the risks of artificial intelligence and machine learning. An email classifier that decides which desk sees a client's instruction looks a lot like a model that needs an owner, a validation record and monitoring. A coding assistant raises a different question: whether AI-written code goes through the same review gate as human code. Sibos 2026 showed the industry's default answer for payments, where agents propose and a person still releases.
The vendor concentration question matters too. Two vendors means two third-party risk files, two exit plans and two sets of data terms. Britain's critical third parties regime covers cloud providers but not the model layer inside them, so a bank has to map that dependency itself.
The Case for Staying on One Vendor
The strongest argument against copying Barclays is integration cost. Every additional vendor brings another identity integration, another data processing agreement, another security review and another admin console. Microsoft's seat now includes Sonnet and Opus within limits, and GitHub Copilot runs on the same Microsoft contract and identity plane. For an organisation with a few hundred engineers, one vendor is often the right call. Our 500-seat coding assistant comparison came down on Copilot Business for that reason.
Barclays' scale changes the maths. With tens of thousands of staff and a legacy estate to modernise, the cost of a second contract is small next to the cost of the wrong tool on a multi-year modernisation programme. The template fits large estates; a 2,000-person company should run the integration maths first.
What to Do Before You Copy the Barclays Split
This Week:
- List every AI workload you run or plan to run, and mark each one as seat-shaped (general assistance, no natural unit) or workload-shaped (emails, tickets, pull requests, documents). Only the second column is a candidate for a second vendor.
- Pull your Copilot weekly active users, not assigned seats, from the Microsoft 365 admin center. If you cannot say how many people used it last week, fix that before you buy anything else.
This Month:
- For any coding assistant pilot, agree three output metrics with engineering before rollout: change failure rate, cycle time from first commit to merge, and post-merge defect rate on AI-assisted changes.
- For any classification or routing workload, write down the misroute rate you accept today from people, then set the same threshold for the model and sample its decisions weekly.
- Ask your model risk team whether an LLM classifier falls inside your SS1/23 model inventory. Get the answer in writing before the pilot goes live.
Before Renewal:
- Price your Copilot renewal against actual weekly use, and move any workload-shaped jobs you identified to a per-unit comparison against a direct model contract.
- Require the vendor on any second contract to support per-user spend caps and an analytics export that works with your data retention settings, since some metrics disappear when zero data retention is on.
The Bottom Line
Barclays has shown that a large regulated enterprise can run a desktop AI vendor and a workload AI vendor side by side, and it has given its rollout a public deadline: half of its developers on Claude Code within three months. That is the useful part of the template. The missing part is the evidence, because the bank has not published a single number about what Claude or Copilot changed. Barclays Private Bank's July outlook described the same pattern across the economy, with model capability moving faster than measured benefit. You can split your vendors the same way Barclays did. Write the output metric into the contract first.
Continue Reading
- Microsoft's New Copilot Bills Autopilot Outside the $30 Seat
- M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Best AI Coding Assistant at 500 Seats: Buy Copilot Business
- True Cost of a Copilot Seat: The Licence Is the Floor
- DBS Put 1,500 Bankers on Agents. The 30% Is a Goal.
- Cursor Alternatives That Survive Security Review Start With Copilot
