AI for government leaders: questions before a pilot

Before you approve an AI pilot as a senior leader, get answers to eight questions: what problem it solves, whether the use is one to avoid, where a person checks the output, who is accountable, what data it uses, how you will know it works, how you would leave the supplier and what you will publish. Each one draws on the AI Playbook for the UK Government, which the Government Digital Service (GDS) published on 10th February 2025.

The playbook sets out ten principles for civil servants and is a long read. These eight are the questions we think a sponsor has to ask in person, and none of them calls for technical knowledge. If you would like more grounding first, our one-day AI for Government Leaders programme is for you, and you need no coding or data science to follow it.

1. What problem are we solving, and for whom?

GDS says the choice of where to use AI must be led by what the business and its users need, and never by what the technology can do. A good answer names the users, the problem and a measure with today’s figure. GDS also asks you to stay open to the conclusion that established technology would solve the problem more easily. Ask the team what else they considered.

2. Is this a use we should avoid?

The playbook lists uses that are not appropriate for the public sector. GDS tells teams to be cautious about fully automated decisions on significant matters, such as someone’s health or safety. It also says AI should not be used on its own in high-risk areas where it could harm the environment or a person’s health, safety or fundamental rights. If the pilot touches decisions about individuals, ask the team to show you where it sits against that list.

3. Where does a person check the output?

Principle 4 is that people keep meaningful control at the right stages. For a high-risk decision, a person validates what the AI produced. Where nobody can review in real time, as with a chatbot, the control has to come at other stages, through full testing before launch and regular checks afterwards. Ask how a user reports a wrong answer and who reads the report.

4. Who is accountable?

GDS tells organisations to nominate a senior responsible owner who is accountable for the use of AI in a specific project. It also recommends an AI governance board, or AI expertise on an existing board, to review and authorise uses. If you are being asked to approve the pilot, find out whether you are that owner and which board has seen it.

5. What data will it use, and is that lawful and secure?

GDS says a team has to carry out a data protection impact assessment before implementing an AI solution, and should take data protection advice early. On security, GDS names threats that are particular to AI. One is prompt injection, where someone hides instructions in the text an AI system reads. GDS also requires services to follow the government’s Secure by Design principles. Ask to see the assessment, and ask who in your security team has reviewed the design.

6. How will we know it works?

GDS states that AI systems are not guaranteed to be accurate and that teams should have a process for testing them. So agree three things before the pilot starts: the measure, the date you will look at it and the result that would make you stop. Write them down. Without a stopping rule, you will be asked to extend the pilot with nothing to judge it against.

7. How would we leave the supplier?

Principle 8 is to work with commercial colleagues from the start. In its advice on buying AI, GDS tells teams to begin with the problem statement and to think about how to avoid being locked in to one vendor. Ask what it would take to move to another supplier in a year. Ask too what happens to the service if the AI part is switched off, because GDS expects fallback processes for that case.

8. What will we publish, and do we need extra approval?

The Algorithmic Transparency Recording Standard is mandatory for all government departments and for arm’s-length bodies that deliver public or frontline services or deal directly with the public. Under it, an organisation publishes how and why it uses algorithmic tools. Ask whether the pilot is in scope. GDS also says any automated response to the public should be clearly identified as one.

Then check approval. Since 1st April 2026 departments approve most of their own digital spending. Even so, GDS gives using AI without internal expertise as an example of a “novel” proposal, and novel proposals need HM Treasury approval. Your department’s assurers will also check the pilot against the playbook’s principles. Our guide to agile business cases and spend controls covers the approval route.

What to do before you sign

Ask the team for one page that answers the eight questions, and go through it with them in person. Put the review date from question 6 in your diary. If the answer to question 4 is you, read the ten principles in the playbook before the pilot starts.