AI in Government: The Problem Is Data, Not the Tool

AI tools are now easy to buy. Whether AI in government works depends on data you can explain, trust and lawfully use. A data-first view of responsible AI.

What you need to know in 60 seconds

Buying an AI tool is now the easy part of AI in government. Departments can pilot almost anything. What decides whether a tool scales, and whether it can be defended in public, is the data underneath it.

GDS reached the same view after the Prime Minister's AI Exemplars programme: departments pilot AI successfully, but scaling beyond pilots proved slow and resource intensive, held back by system-level barriers that include data readiness[1]. GDS now argues that any AI readiness assessment should start with data maturity[2].

This piece sets out what a weak data foundation looks like from the inside, why responsible AI is mostly a data discipline, and five checks to run before buying the next tool.

Why is the tool now the easy part of government AI?

A few years ago the hard part of AI in government was getting access to a capable model. Today that is a procurement exercise. The AI Playbook for the UK Government assumes teams will choose between many options and asks them to use the right tool for the job[3].

The constraint has moved. The State of Digital Government Review counted over 140 AI use cases already identified across central government, yet it reported that only 8% of AI projects showed measurable benefits[4]. GDS's lessons from the AI Exemplars point the same way: similar work was repeatedly being done in parallel across departments, and scaling was most often held up by barriers no single tool could remove[1].

A better tool fixes none of that. A better foundation does.

What does a weak data foundation look like inside a department?

It rarely looks like a crisis. It looks like ordinary friction that everyone has learned to work around:

The review cites central analysis of AI adoption planning across the public sector that highlighted data accessibility and quality as significant obstacles[4]. Put an AI tool on top of that foundation and it will give fluent, confident answers from data nobody can vouch for.

Why is responsible AI a data problem first?

Most of what responsible AI asks of a public body comes back to data.

Accuracy depends on data. The AI Playbook is blunt: AI systems heavily rely on the quality and quantity of data, and poor quality, biased or noisy data can lead to inaccurate outcomes[3]. Bias is usually inherited from the records a model learns from or retrieves, not invented by the model.

Lawfulness depends on data. Government's guidance on AI-ready datasets lists ownership, licensing and a lawful basis for computational use among the legal needs, and asks whether provenance is known: where the data comes from and how it is produced[5].

Transparency depends on data. Central government departments, and arm's length bodies within scope, are required to use the Algorithmic Transparency Recording Standard[3], which means explaining in public how an algorithmic tool works and how it is used. That is very hard to do honestly for a tool fed by data nobody can trace.

GDS put it most simply: poor quality data produces unreliable AI outputs, and good data that is poorly understood or poorly governed cannot be considered AI ready either[2].

What should a department check before buying another AI tool?

Five checks, each answerable within a week by people who already work in the department:

  1. Provenance. Can you trace the data the tool will use back to its source systems and every step that changed it?
  2. Lawful basis. Has someone confirmed the data can lawfully be used for this purpose, and not only the purpose it was collected for?
  3. Ownership. Is there a named owner who answers for this data's quality and has the authority to fix it?
  4. Measured quality. Do you know how complete, accurate and current the data is, as numbers rather than impressions?
  5. Explanation. When the tool gets something wrong, could you show a citizen, a journalist or a court why?

If most of those get a confident yes, the tool is probably worth piloting. If not, the money is better spent on the data first, and the pilot will move faster afterwards.

What does data-first AI look like in practice?

It means building the foundation with the AI use in view, instead of running a separate data programme that never quite finishes.

When the Infected Blood Compensation Authority was set up, Mayfair helped build its Azure data platform from nothing, with a master data catalogue, role-based access and a Zero Trust blueprint from the start. All of the authority's case data was brought into one governed place, and the platform was certified as AI-ready, paving the way for automated document triage in a later phase. The AI came second, and that is why it could come at all. Read the IBCA case study.

At the Office for National Statistics, the work was about making sensitive data usable without exposing anyone: privacy-preserving anonymisation and automated disclosure checks on every dataset entering the Integrated Data Service. That is the unglamorous foundation that lets analysts, and in time AI tools, work across departmental data safely.

GDS is taking the same approach centrally. Working with The National Archives, it used data maturity as the framework for judging AI readiness before choosing any technology, and found the archives' legal data close to AI ready because both the data and the organisation around it met the bar[2].

Common questions about AI and data in government

What makes government data ready for AI?

Data is ready for AI when it is well managed, understood and governed: you know where it came from, how good it is, who owns it, and that you can lawfully use it for the purpose in hand. Government's guidance groups this into four areas: technical optimisation, data and metadata quality, organisation and infrastructure context, and legal, security and ethical compliance.

Does data have to be perfect before a department uses AI?

No. It has to be good enough for the specific use, and you have to know how good it is. A tool that drafts internal summaries can tolerate more than one that informs decisions about individuals. The aim is to measure quality, not to wait for perfection.

What is the Algorithmic Transparency Recording Standard?

It is the government standard for publishing information about the algorithmic tools public bodies use. The AI Playbook says central government departments, and arm's length bodies within scope, are required to use it, and other public bodies are encouraged to.

How does data governance support responsible AI?

Governance supplies what responsible AI depends on: named owners, agreed definitions, quality measures, access controls and a record of lineage. Without them, commitments on fairness and transparency can't be checked.

Where should a department start?

Start with one AI use you actually want, and assess the data behind it rather than the tools on offer. Run the five checks above, measure the quality of the main datasets, and fix the gaps that matter for that use. You end up with a better pilot and a reusable piece of the foundation.

Mayfair IT Consultancy is a data-first government AI consultancy working across UK central government, and our data governance team builds the ownership, quality and lineage that responsible AI depends on. If you'd like a second opinion on whether your data is ready for the tool you have in mind, talk to us.

Related reading