Government Digital Services - AI

We carried out user research to find out whether the Service Manual meets the needs of teams building services that use AI.

• Teams find it hard to keep up with how quickly AI changes, and aren't confident they understand it well enough to make good decisions.
• Teams can't easily tell whether they're following good practice, because the guidance doesn't show enough concrete examples.

The guidance covers the right topics, but teams wanted more practical, applied detail to help them act on it. The research surfaced four areas where teams wanted the guidance to go further.

• Monitoring AI. Teams wanted clearer guidance on monitoring AI in a live service: both how to do it and what to watch for. They were unsure what to track, how to tell whether the AI was still performing as intended and what to do when it was not.
• Accuracy. Teams wanted practical guidance on how to keep AI outputs accurate and how to maintain reliability as the service and the underlying technology change.
• Bias. Teams wanted to know how it arises, how to test for it and how to monitor for biased or unfair outcomes.
• Security. Teams wanted more detail on engaging and working with security professionals — including when to involve them, what to ask for, and how to build that relationship into delivery rather than treating it as a one-off check.

• Worked with stakeholders across government to get answers to what the users wanted to know.
• Researched AI in depth over several months — the technology and how teams apply it in services.
• Designed and rewrote the guidance, restructuring it around the gaps the research identified.