Summary

• HMRC is recruiting a deputy director level Head of AI Assurance on a salary of £86,000 to £110,000.
• The role joins an existing Chief AI Office responsible for governance, risk, transparency, and human oversight.
• Permanent assurance functions are becoming necessary as government departments move AI into service delivery, compliance, and workforce systems.

HM Revenue & Customs is creating a senior leadership post devoted to AI assurance, adding permanent governance capacity as artificial intelligence moves deeper into tax administration, compliance work, employee tools, and digital services.

HM Revenue & Customs is recruiting a deputy director level Head of AI Assurance in Manchester, with a salary of between £86,000 and £110,000 and applications open until 14 September. The role is intended to oversee the safe and effective use of AI across a department whose systems handle some of the most sensitive financial and personal information in government.

HMRC is not building that function from scratch because it already has a Chief AI Office covering strategy, governance, risk, transparency, and human oversight. The department’s published organisation structure makes service and system owners responsible for decisions made with AI, while the central AI function establishes standards and assurance around how the technology is used.

That separation prevents assurance from becoming a mechanism through which operating teams transfer responsibility to technical specialists. A central function can test controls, provide methods, and challenge a proposed deployment, yet managers running a tax, compliance, or customer service process still own the consequences of the system once it enters production.

Production AI requires repeatable controls

Government departments have spent several years publishing principles, establishing pilots, and experimenting with generative tools, although permanent deployment creates a different workload. Systems need evaluation before launch, monitoring afterwards, clear ownership when models change, defined routes for incident handling, and evidence that human review operates as designed.

HMRC’s environment raises the threshold because not every use case carries the same consequences. An internal tool that summarises guidance presents a different risk from a system that helps prioritise compliance cases, influences communication with taxpayers, or contributes to a process involving debt or fraud.

Assurance therefore has to vary with the decision being supported rather than applying a single checklist to every AI product. Technical accuracy may be only one concern: teams also need to understand the quality of underlying data, possible bias, whether outputs can be challenged, how supplier changes affect performance, and what happens when the service or model becomes unavailable.

HMRC’s 2026 transformation roadmap already places AI inside a broader digital programme rather than treating it as a standalone innovation exercise. The department has been expanding digital customer interaction and considering AI across compliance and productivity work, which creates demand for governance that can operate continuously rather than convening only around unusual projects.

Assurance is becoming part of public sector procurement

The same development will affect suppliers because government AI systems depend on cloud providers, software vendors, model developers, consultancies, and systems integrators. Departments will increasingly need contractual evidence around model evaluation, logging, access control, monitoring, incident investigation, data handling, and changes introduced after deployment.

Traditional cyber security and data protection remain essential, but neither fully answers whether an AI system is reliable enough for a particular task. A model can be securely hosted and lawfully process data while still making too many errors, behaving differently after an update, or being used by staff in situations for which it was never evaluated.

Operational design therefore sits alongside technical assurance. Employees need enough time and knowledge to review outputs, managers need routes for escalating failures, and affected taxpayers need processes through which incorrect decisions can be challenged. Adding a nominal human approval step achieves little if the surrounding workload encourages employees to accept machine output automatically.

Cross government work has already identified evaluation, safety, legal questions, commercial arrangements, and adoption planning as recurring needs across AI projects, suggesting that specialist assurance teams will become shared organisational infrastructure rather than temporary support for individual pilots.

Creating a deputy director post does not prove that HMRC has solved those problems, but it establishes a clearer point of responsibility as the technology becomes more deeply embedded. The eventual measure of the role will be whether useful systems reach production without unnecessary delay, high risk deployments are challenged early, and failures can be traced through an assurance process that existed before something went wrong.