HMRC scales Microsoft Copilot to 28,000 staff with 50,000 target

TL;DR:

• HMRC is rolling out Microsoft 365 Copilot to 28,000 staff initially, with plans to extend to 50,000 of the department’s roughly 64,000 employees — one of the largest Copilot deployments in UK government.
• James Mitton, the department’s first Chief AI Officer, frames the programme as transforming a tax authority serving more than 40 million taxpayers, with around 20,000 customer-support staff in scope for AI-augmented workflows.
• An AI complaint-handling agent is in development to summarise customer complaints and guide handlers — a concrete agentic pilot rather than a generic productivity drive.

The rollout sits inside HMRC’s Transformation Roadmap, which Mitton’s team describes as “an agile department supported by a modern IT infrastructure, with innovation driving a better customer experience”. Exchequer Secretary Dan Tomlinson MP is positioning Treasury investment in the programme as core to closing the tax gap.

Why scale matters here

A 50,000-seat Copilot deployment is meaningful for two reasons. First, on the supply side: this is one of the largest public-sector Copilot contracts in the UK and a reference point Microsoft will use in NHS, MoD and devolved-government conversations over the next 18 months. Second, on the policy side: HMRC handles compliance and enforcement workflows where AI-assisted decisions carry direct citizen-impact risk. The department’s framing — “any decisions will be reviewed and signed off by a human in the process with the relevant tax expertise” — places explicit human-in-the-loop guardrails on the deployment, in line with HMRC’s “AI assurance, ethics and risk management framework, tested with external ethics experts”.

The complaints agent is the interesting bit

The complaint-handling agent is a sharper signal than the headline seat count. HMRC has used machine learning for “compliance targeting, document analysis, debt risk prediction, and customer contact handling” for years, but Matt Vick, Head of Futures and Innovation, frames the agentic step as a discontinuity — moving from rule-based ML to systems that respond to natural-language prompts, identify what information is needed next, and structure responses around customer focus. That is a different operating model, and a more sensitive one to get wrong: a botched ML model produces poor analytics; a botched agent gives a citizen the wrong tax answer.

Context against the data-readiness gap

This rollout lands in the same week as the Aker Systems paper warning that only 14% of public-sector bodies have AI-ready data platforms. HMRC is on the more advanced side of that curve — its years of ML deployment imply some data infrastructure exists — but at 64,000 staff, even small failure rates in the agent-driven workflows would touch millions of citizen interactions. Compared to recent UK government experimentation surveys (65% of staff trying AI but only 30% with integrated workflows), HMRC’s structured Copilot programme is closer to integration than experimentation.

Looking forward

Watch for the complaints agent’s pilot results: HMRC’s framework requires human sign-off, but the productivity numbers will determine whether agentic systems extend to the 20,000 customer-support team. The department’s quote — “what started as a conversation on the day has grown real momentum since” — describes the cultural shift Microsoft will be selling to NHS, DWP and local authorities over the next year as Microsoft pushes Copilot into other large public-sector estates.