Scaling AI in the public sector: appropriate transparency and governance
Government is moving from pilots to delivery. The 2025 AI Opportunities Action Plan is being executed, with the 2026 review reporting 38 of 50 actions complete. Foundations are coming online: priority datasets via the National Data Library and public sector compute through Isambard AI. Meanwhile, teams are deploying practical tools like AI scribes for meetings and systems that extract information from planning documents.
The AI Playbook: what it expects
The Government Digital Service's AI Playbook sets 10 principles that keep programmes lawful, useful and safe. It has been updated since 2025 to reflect new terms, security risks and public sector examples. Two themes do the heavy lifting as adoption grows: appropriate transparency and strong governance.
Appropriate transparency
There is no blanket duty to publish every use of AI. Still, public bodies are expected to provide meaningful transparency, especially where tools influence decisions that affect people. The Algorithmic Transparency Recording Standard (ATRS) is mandatory for central government departments and relevant arm's length bodies when a tool significantly shapes decisions with public effect or directly interacts with the public. Since the first ATRS record in July 2022, there are 125 records across a wide range of use cases, and that number will rise.
Where transparency is moving
- Procurement and contracts: PPN 02/24 requires you to ask suppliers how they use AI in tenders and service delivery. Expect contracts to include obligations on disclosure, data sources, model changes, human oversight, logging, incident reporting and audit access.
- Market standards: From 2 August 2026, key EU AI Act transparency duties apply in the EU, including marking AI-generated content and adding certain high-risk systems to a public register. The European Commission's draft code on content labelling (December 2025) points the same way. The Act does not bind UK public bodies directly, but suppliers operating in the EU will bring these practices with them.
- Legal pressure: In 2025, the First-tier Tribunal required HMRC to confirm whether and how AI was used to refuse some R&D tax credit claims. The point was to reduce public concern and support informed debate. Expect case-by-case scrutiny of what "enough transparency" looks like.
Practical transparency actions to start now
- Use the ATRS by default for any tool that influences decisions or interacts with the public, and keep entries current.
- Publish short, plain-language service notes for significant AI-enabled services: purpose, data used, safeguards, human review, appeal routes and how to contact a human.
- Adopt a policy for labelling AI-generated or AI-assisted content where relevant, and make it consistent across channels.
- Build "FOI-ready" documentation: decision logs, model/version history, evaluation results and risk assessments.
- Map where suppliers use AI on your behalf and mirror your transparency standards in their contracts.
Governance
The Playbook is clear: successful AI programmes need strong governance. That can mean a new AI governance process or updates to existing ones, but decision rights, accountability and assurance must be explicit.
Build the right structures
- Set up an AI governance board or embed AI expertise into an existing board. Give it clear decision rights over use cases, risk thresholds, go/no-go gates and incident handling.
- Refresh risk management: define AI risk tiers, pre-deployment checks, red-teaming where proportionate, and a live risk register tied to service owners.
- Establish AI quality assurance: test for accuracy, security, reliability, drift, bias and privacy before and after deployment, and publish evaluation summaries where appropriate.
- Engage legal, compliance and data protection teams early, including during product development, not just at the point of launch.
Maintain an AI systems inventory
Keep a single inventory listing each AI system's purpose, owner, data sources, models, integrations and current status. This underpins risk control, auditability and accountability. As usage scales across more vendors and data sources, move from static spreadsheets to dynamic registers with automation.
- Integrate the inventory with procurement, architecture reviews and data protection impact assessments so nothing bypasses oversight.
- Tag sensitive data and permissions; log model and prompt changes; record human-in-the-loop controls and appeal routes.
- Require suppliers to attest to changes, incidents and significant model updates, and align on monitoring metrics.
Roles and capability
Government is adding senior expertise (for example, AI fellows) and delivery teams such as the Incubator for AI. Expect further changes as usage grows. Make responsibilities crystal clear and invest in skills for policy, commercial, digital and frontline teams.
- Define RACI across business owners, product managers, SIRO, DPO, CISO, delivery leads and an ethics or assurance lead.
- Provide targeted training for policymakers and commercial teams making AI decisions. See the AI Learning Path for Policy Makers.
- Create a cross-government community of practice to share patterns, evaluation methods and procurement clauses.
90-day checklist
- Stand up (or refresh) your AI governance board and approve decision rights, thresholds and escalation paths.
- Complete or update your AI systems inventory and link it to procurement and DPIA processes.
- Implement ATRS for relevant services and publish at least one plain-language service note for a high-impact use case.
- Adopt a supplier AI transparency clause set and bake it into all new procurements and change controls.
- Set baseline monitoring: accuracy, bias, security and drift, with reporting to service owners and the governance board.
What this adds up to
There is no single blueprint for "good" transparency or governance. The right answer depends on the service, the data and the people affected. Start early, keep it practical, and make it auditable. That is how you stay lawful, keep trust and scale AI with confidence.
References
- Artificial Intelligence Playbook for the UK Government - GOV.UK
- PPN 02/24 Improving Transparency of AI use in Procurement - GOV.UK
Your membership also unlocks:
