Half of Barclays' developers on Claude Code by December

The billable hour that kept a bank's legacy code alive. Barclays just put a deadline on it.

Barclays and Anthropic announced an expanded collaboration on October 1, 2026. The scope covers software development, legacy modernization and operational workflows.

The commitment is a number with a date. Claude Code adoption reaches 50% of Barclays' developer population by the end of 2026.

A majority of software engineers follows in 2027. That is a staffing plan written as a technology announcement.

This is not a first contact. A colleague knowledge assistant has run on the models since 2025.

The markets business already uses them on client inquiries. That platform handles roughly 120,000 emails a day.

Group Co-Chief Operating Officer Craig Bright named the target. The bank is using Claude to "modernize legacy platforms, improve software quality".

His co-COO Anne Marie Darling framed it around routine work. Fewer manual steps, more capacity for complex problems.

The commercial context sits outside the announcement. Barclays runs a multi-year efficiency programme targeting about £2 billion in savings.

AI contributes part of that figure. The bank has not said how much.

Start with what legacy modernization has always been. It is the largest line in a big bank's technology budget and the one that never finishes.

Banks do not do this work themselves. They buy it by the hour from systems integrators and offshore development firms.

That market exists because the work is unglamorous, enormous and slow. Thirty years of accumulated code, poorly documented, written by people who left.

Barclays has now bought a different thing. Not hours, a licence.

The second concession is sharper and less discussed. Modernizing a core banking platform means giving a model the source code.

That is the most sensitive asset a bank holds. Payment logic, limit checks, account structures, the rules the regulator examines.

Barclays says governance, security controls and human oversight apply. It has not published the controls.

The third concession is strategic. Technical debt was the standing excuse for why banks cannot move quickly.

Remove it and the excuse goes too. A bank that can rewrite its own plumbing in quarters rather than decades has nowhere left to hide behind.

Read the structure of the target. Fifty percent by December, majority in 2027.

Nobody publishes an adoption percentage for a pilot. Percentages are for headcount planning.

Now read what is missing. Barclays quantified no reduction in development effort and no reduction in support time.

That silence is the most informative part of the announcement. A bank with a measured productivity gain publishes it.

A bank with a target and no measurement is committing before the evidence arrives. That is a reasonable thing to do and it should be described accurately.

The peer pattern says the same. HSBC opened authorised banking data to clients' AI tools in the same week.

BNP Paribas signed a five-year agentic deal with Google Cloud on September 24. DBS scaled credit agents to 1,500 bankers in August.

Four large banks, four vendors, one quarter. None of them has published an error rate.

Here is what would settle this, and nobody has it.

A developer productivity figure needs a baseline, a comparison group and a measure of defects. Lines of code and pull request counts measure activity, not output.

Legacy modernization has a harder problem still. The failure mode is not slow delivery, it is a rewrite that behaves differently in an edge case nobody tested.

Banks discover those edge cases at settlement, in a reconciliation break, or in a regulatory exam. The lag between writing the code and learning it was wrong runs to years.

Supervisors have noticed. The Australian securities regulator plans to examine customer impacts from banks' AI applications.

The Basel Committee is reviewing how existing operational loss categories capture AI risk. Neither has concluded.

The UK's critical third parties regime is the one that binds Barclays. A model doing load-bearing work on core systems is exactly what that regime was written for.

The systems integrator first, and it sits outside the building. Legacy maintenance billed by the hour is the business this attacks directly.

Offshore development contracts are priced on headcount. A tool that cuts the headcount cuts the contract.

Then the graduate intake. Junior engineering work on a legacy platform is reading old code and making small changes.

That is precisely what Claude Code does first. The bottom of the ladder is always the first rung to go.

Then the mainframe specialist. Scarcity was the entire value of knowing a forty-year-old system.

A model that reads the code removes the scarcity, not the system. The expert becomes a reviewer rather than the only route in.

Then the change-the-bank project office. Fewer, shorter projects means a smaller apparatus managing them.

No cuts were announced here. The £2 billion efficiency programme is where that announcement already lives.

Watch for the first published productivity figure. Barclays committing to a percentage without one is the gap to track.

Watch the 2027 target harder. Fifty percent is adoption, and a majority of engineers is dependence.

Territory: +machines on the build, +incumbents on the bank.

Legacy maintenance billed by the hour is the pool compressing, while the code and the accountability stay inside Barclays.