AI in the public sector: what's really happening, and what actually works
Our latest Spotlight webinar brought together two very different perspectives on AI in government: a researcher who has mapped how local authorities are using it, and a regulator who has put it to work on real consultation data. Here's what we learned.
Artificial intelligence is the topic that won't leave the public sector's agenda alone. But between the hype and the headlines, it can be hard to get a clear picture of what's actually happening on the ground. That's what this Spotlight session set out to do, with two speakers who could each answer the question from first-hand experience: Kevin Fenning of Evidence First, and Adrian Price of the Information Commissioner's Office (ICO).
If you missed it, or want to watch it again, the full recording is available here.
The view across local government: 101 use cases and counting
Kevin Fenning opened the session with findings from research he carried out for the Local Policy Innovation Partnership (LPIP) Hub, drawing on interviews with 35 people across 22 organisations and a review of 101 real-world AI use cases in local government.
The headline: adoption is maturing fast, but unevenly. While some councils are well into deployment, around half are still at an early, exploratory stage. The most common uses will be familiar to many: customer service chatbots, Copilot-style productivity tools, and note-taking assistants in social care.
But it was the long tail of niche applications that raised eyebrows. Robots that seal cracks in roads before they become potholes. Computer vision that detects seatbelt use. And a project using AI to identify bus shelters that should be paying business rates, which, if scaled across London, could be worth an estimated £10m.
Kevin was candid about the gaps too. Of the 101 use cases reviewed, only 44 came with quantified benefits, and just 14 had been monetised. The evidence base is thinner than the enthusiasm, which makes it harder for councils to build robust business cases and learn from one another.
He also flagged the risks that keep practitioners up at night: procurement lock-in, skills shortages, and quality assurance of AI outputs. Two examples stuck with the audience: an AI note-taking tool that hallucinated a suicide risk into a social care record, and a transcription tool that swapped the genders of the people in a conversation. Both are powerful reminders of why human oversight isn't optional.
Looking ahead, Kevin's advice was practical: keep contracts flexible to avoid lock-in, make sure AI policies keep pace with developments like agentic AI, and be ready for the "double-edged sword" of residents using AI tools themselves, whether that's drafting a complaint or generating hundreds of consultation responses.
You can read Kevin's full report here
The view from inside a regulator: the ICO's honest verdict on Copilot
Adrian Price, who leads regulatory policy projects at the ICO (and is a Citizen Space user), gave a refreshingly honest account of what happened when the ICO rolled out Microsoft Copilot to around 1,000 staff.
Rather than a glossy success story, Adrian offered a practitioner's scorecard. On consultation analysis, Copilot proved genuinely useful for:
- Identifying themes across sets of free-text responses
- Answering specific questions about what respondents said
- Checking terminology, such as where a draft says "must" versus "should" versus "could"
- Producing meeting minutes from Teams transcripts
And it was noticeably weaker at:
- Quantitative analysis, where results simply couldn't be trusted
- Applying a style guide (the ICO is trialling a dedicated tool, PerfectIt, for that instead)
- Weighting responses by relevance or expertise
- Tagging and categorising free text, where it would sometimes get things exactly backwards
Out of that experience came the ICO's rules of thumb, which any team could adopt tomorrow: keep datasets small, keep queries specific, set ground rules (like requiring the AI to reference the responses it draws on), and keep humans firmly in charge of quality assurance.
One benefit Adrian highlighted was less tangible but no less real: confidence. Analysts using AI as a first pass felt more assured that they hadn't missed something in a large body of responses, even though the human analysis remained the authoritative version.
The thread running through both talks
Two speakers, two very different vantage points, one shared conclusion: AI is a genuinely useful assistant and a genuinely unreliable authority. The organisations getting value from it are the ones treating it as a first pass, not a final answer, with human judgement doing the quality assurance throughout.
That's a principle we share at Delib. Citizen Space's data API lets you connect your consultation data securely with the analysis and AI tools of your choice, so your team can work faster while staying firmly in control of the findings.
Join us next time
Make sure you're signed up to our email list to be the first to hear about our events and webinars.