The UK Public Sector AI Adoption Gap
Why UK public sector AI stalls between pilot and value
Nearly every public sector organisation in the UK is now using AI in some form. Very few can show what it changed.
This report brings together the published evidence from government trials, independent research, and industry analysis. It explains why so many pilots stop short of value, and what the organisations getting it right do differently.
Every figure here is named, dated, and linked. Where the evidence is thin, we say so.
Executive summary
Everyone is adopting. Almost nobody is scaling.
Two years ago the public sector question was whether to use AI at all. That question has been settled by staff, not by strategy. Adoption arrived from the bottom up, through browsers and free tools, faster than any technology before it.
The numbers show how complete that shift is.
-
95% of responding councils are using or exploring AI Source: LGA, 2025
-
79% of charities use AI tools in day to day work Source: Charity Digital Skills Report, 2026
-
41% of public sector working time could be supported by generative AI Source: Alan Turing Institute, ONS, and DfT, 2025
Now look at the second set of numbers, from the same surveys.
-
7%
of those councils rate themselves as leaders in their use of AI
- 7%
- rate themselves as leaders
- 95%
- are using or exploring AI
Source: LGA, 2025
-
38%
of charities use AI actively or strategically
- 38%
- use AI actively or strategically
- 79%
- use AI tools day to day
Source: Charity Digital Skills Report, 2026
-
4%
of teachers report using generative AI for marking
- 4%
- of teachers use generative AI for marking
- 49%
- of education working time AI could support
Source: Alan Turing Institute, 2025
That is the adoption gap. Near-universal use sitting on top of very thin claims to maturity. The tools are in the building. The value is not yet on the balance sheet.
The gap is not caused by the technology. Three government departments ran trials of the same product and reached three different conclusions, which we cover in Chapter 03. What separated them was scope, use case selection, readiness, and how they chose to measure.
Closing the gap takes three things, in this order.
-
Step 1: Direction
Decide what problem you are solving, and for whom, before you decide which technology solves it. Most stalled pilots reversed that order.
-
Step 2: Proof
Build something real, with success measures agreed before the build starts. A pilot with no baseline cannot produce evidence, only opinion.
-
Step 3: Trust
Design security, governance, and accountability in at the start. Added afterwards, they stop a working pilot from ever reaching production.
Chapter 01
Data and sources
This report contains no Phoenix survey and no Phoenix telemetry. It draws entirely on published evidence, and it names all of it. There are three groups of sources.
-
Government evaluations
Published findings from the Government Digital Service, the Department for Work and Pensions, the Department for Business and Trade, the Department for Science, Innovation and Technology, and the National Audit Office. These are the most useful sources available on public sector AI, because the organisations concerned had no product to sell.
-
Independent research
Work from the Alan Turing Institute with the Office for National Statistics and the Department for Transport, the Local Government Association, the Charity Digital Skills Report, and academic research into council AI readiness led by Heriot-Watt University.
-
Industry analysis
Forecasts and analysis from Gartner, and the widely cited MIT study on enterprise generative AI. We flag where that work is preliminary.
-
What we do not have
Phoenix does not yet publish outcome data for AI Surveyor engagements. We are not going to borrow someone else’s numbers and imply they are ours. When we have our own results, with customer permission, they will appear here and be dated.
Chapter 02
The prize is real, and it has been measured
Start with the size of the opportunity, because it is genuinely large and it has been quantified by people with no commercial interest in the answer.
Researchers at the Alan Turing Institute, working with the Office for National Statistics and the Department for Transport, mapped what public sector staff spend their time doing. They then compared those activities against what generative AI can currently do.
41%
Around 41% of public sector working time is spent on activities that generative AI could support. The figure varies by sector, and by how close to the front line the role sits.
The government has put a figure on the wider prize too. Its State of Digital Government Review found more than £45 billion a year in unrealised savings and productivity benefits, worth 4 to 7% of public sector spend, available through full digitisation of public services.
We should be straight about that number. The Commons Science, Innovation and Technology Committee later described it as worryingly optimistic, and the former head of the government’s Incubator for AI called it a reasonable guess that is very hard to measure. Treat it as a direction of travel, not a business case.
Where trials have measured actual time, the results are more modest and more useful. The Government Digital Service ran the largest deployment of Microsoft 365 Copilot in any organisation to date: 20,000 staff across 12 government organisations, over three months.
-
26 minutes
saved per person, per day
-
13 days
a year, if sustained across a full working year
-
82%
of users would not want to go back to working without it
Source: Government Digital Service, M365 Copilot experiment findings, published June 2025.
More than 70% of those users agreed that Copilot cut the time they spent searching for information and doing mundane tasks. Document drafting saved 24 minutes on average. Building a presentation saved 19.
The question is no longer whether AI can help. It is whether your organisation can prove that it did.
Chapter 03
Same tool, three departments, three answers
This is the finding that should change how public sector organisations plan AI work.
Three UK government departments ran evaluations of the same product, Microsoft 365 Copilot, over overlapping periods. All three published. They did not agree.
Same licence. Three verdicts.
-
26 min
a day saved. Government Digital Service, 20,000 staff, 12 organisations
Source: GDS, published June 2025
-
19 min
a day saved, measured against a comparison group. DWP, 3,549 licences
Source: DWP, published January 2026
-
None
No robust evidence that time savings led to improved productivity. DBT, 1,000 licences
Source: DBT, published August 2025
The Department for Work and Pensions ran the most rigorous of the three. It compared 1,716 licence holders against 2,535 non-users, added interviews, and adjusted for differences between the groups. It found 19 minutes a day, plus small but statistically significant improvements in perceived work quality and job satisfaction. Neurodivergent staff reported particular benefit.
The Department for Business and Trade found something different. Staff liked the tool, with 72% satisfied or very satisfied. But the department could not find robust evidence that time savings had turned into productivity. Some tasks got slower. Slide production was around seven minutes faster but needed correcting for quality. Analysis in Excel took longer and was less accurate.
Usage was the most revealing number
1.14
Copilot actions per day, per average participant, across 63 working days.
Read that again. Roughly one action a day. This was not a technology that failed to deliver. It was a technology that was barely used.
The department was also honest about its own method. Productivity was not a key aim of the evaluation, so limited data was collected on it. The negative result is partly a measurement design story.
What differed
The licence was identical in all three cases. What varied was everything around it.
-
Whether specific use cases were identified before rollout, or licences were handed out and left to find their own purpose.
-
Whether a baseline existed, so that any change could be measured against something.
-
Whether staff were trained and supported, or expected to work it out.
-
Whether the evaluation was designed to detect productivity at all.
-
Whether security and data handling concerns were resolved up front. GDS found that unresolved concerns reduced the benefit for a minority of its users.
Nothing in these three evaluations is a technology failure. Every difference between them is a decision that was made, or not made, before the work began.
Chapter 04
Five reasons pilots stall
The pattern is consistent across the evidence. Pilots stop for the same five reasons, and none of them is model performance.
No decision gate
The technology gets chosen before the problem is understood. Gartner reports that at least half of generative AI projects were abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value. Its earlier forecast had been 30%.
A decision gate is simply a point at which you stop and choose, with evidence in front of you. Copilot, Copilot Studio, low-code, pro-code, or machine learning are different answers to different problems. Picking first and discovering later is expensive.
Data foundations that are not ready
Councils responding to the Local Government Association named data availability, quality, and storage as their second largest area of unpreparedness, with 68% saying they were not very or not at all ready. Workforce skills came first at 71%.
The National Audit Office reached the same conclusion about central government. Updating legacy systems and improving data quality and access is fundamental to exploiting AI, and it takes time.
Governance arriving late
When the National Audit Office surveyed 87 government bodies, only 30% had risk and quality assurance processes that explicitly covered AI risks, though a further 46% planned to introduce them. Only 21% had an AI strategy at all.
Governance added after a pilot works is not a formality. It is the reason a working pilot never reaches production.
A legacy estate in the way
The State of Digital Government Review found 28% of central government systems classed as legacy, up from 26% the year before. The picture is worse in parts of the sector that carry the most operational risk.
| Part of the public sector | Reported legacy share of estate |
|---|---|
| Central government | 28% |
| NHS trusts | 10% to 50% |
| Police forces | 10% to 70% |
| Organisations unable to quantify their legacy estate at all | 15% |
Source: State of Digital Government Review, DSIT, January 2025.
Maintaining those systems often costs three to four times what a modern equivalent would. That is money not available for the thing you want to do next.
Feasible is not the same as adopted
This is the most overlooked reason of all. The Turing research found that 49% of education working time sits on activities generative AI could support, the highest of any sector. It also found that only 4% of teachers report using generative AI for marking.
Technical feasibility and successful adoption are different problems. One is solved by procurement. The other is solved by change management, training, permission, and trust.
The wider market signal
The most quoted figure in enterprise AI comes from MIT’s NANDA initiative, which reported that around 95% of enterprise generative AI pilots produced no measurable impact on the bottom line. It is worth being careful with it. The work was preliminary, was not peer reviewed, and its method has been challenged. Treat it as directional.
The direction, though, matches everything above. Very few organisations are converting pilots into value, and the reasons are organisational rather than technical.
Chapter 05
The sector picture
Public sector is not one market. Eleven sectors, eleven sets of pressures, eleven different first moves. Here is where each one sits, and where we would start.
-
Central Government
Where the sector sits
Around 28% of systems are legacy. Spend controls are tight and scrutiny is constant. Three departments have now published Copilot evaluations that disagree with each other.
Where we would start
Pick one process with a named owner and a measured baseline. Publish the result internally, whichever way it goes.
-
Local Government
Where the sector sits
95% of responding councils are using or exploring AI, but only 7% consider themselves leaders. Funding, staff skills, and staff capacity are the three barriers councils name most. Reorganisation is reshaping the map underneath all of it.
Where we would start
Fix the data foundations before building agents. A single view of the resident is worth more than a chatbot on top of fragmented records.
-
Policing
Where the sector sits
Legacy estate runs as high as 70% in some forces. Officer time is the currency that matters. OFFICIAL-SENSITIVE handling is not negotiable.
Where we would start
Target back office and administrative burden first, not frontline decision making. Prove time returned to the watch.
-
Fire and Rescue
Where the sector sits
Reliability, interoperability, and control room integration come before anything else. Capital funding is uneven and estates vary widely between services.
Where we would start
Start where failure is recoverable. Prevention, protection, and reporting workloads, not incident-critical systems.
-
Defence
Where the sector sits
Every claim needs to be provable and every capability accredited. Classification boundaries shape what is possible before any technology conversation starts.
Where we would start
Define the governance and classification envelope first. Scope the use case inside it, not the other way round.
-
Healthcare
Where the sector sits
Around a third of working time sits on activities AI could support, the lowest of any sector, because so much of the work is human. Trusts report legacy estates of anywhere between 10% and 50%.
Where we would start
Administrative and documentation burden, with clinical safety and information governance involved from day one rather than at approval.
-
Charity and Social Impact
Where the sector sits
79% of charities use AI tools in day to day work, but only 38% use them actively or strategically. Accuracy concerns hold back 49%, and data privacy and security concerns hold back another 49%. 60% now have an AI policy or are writing one.
Where we would start
A policy staff can follow, then one funded use case that returns hours to the mission.
-
Housing
Where the sector sits
Regulatory scrutiny is rising alongside tenant expectations and ageing stock. Damp and mould, repairs, and complaint handling all generate high volumes of contact and evidence.
Where we would start
Tenant contact and repairs triage, with a complete audit trail for the regulator.
-
Schools and Trusts
Where the sector sits
Education has the highest AI potential of any sector at 49% of working time. Yet only 4% of teachers report using generative AI for marking. Safeguarding sits above everything.
Where we would start
Teacher workload relief, with safeguarding and data protection designed in and a trust-wide policy behind it.
-
Further Education
Where the sector sits
Multi-site estates, tight funding, and a national skills agenda to deliver. Colleges support learners from 16 to 60 with very different needs.
Where we would start
Timetabling and learner administration, measured in staff hours returned rather than in technology deployed.
-
Higher Education
Where the sector sits
Decisions are federated across IT, academics, finance, and senior leadership. Research data management, student experience, and cyber resilience compete for the same budget.
Where we would start
One faculty or one professional service function. Not an institution-wide programme.
Sector figures sourced in the source register. Where no published figure exists for a sector, the description is qualitative and no number is implied.
Chapter 06
What good looks like
Everything above points the same way. The organisations getting value from AI are not the ones with the best technology. They are the ones that decided what they were solving, proved it at small scale, and built the controls in from the start.
That is what AI Surveyor does.
AI Surveyor is a fixed-price engagement that takes a public sector organisation from ambition to a working, governed pilot. It follows an Advise to Build cycle, with security and governance underneath both rather than bolted on afterwards.
-
Advise. Find the right starting point.
Before any technology is chosen, Surveyor establishes where you are today. That means an AI strategy and readiness session, functional discovery of the process itself, technical discovery of the architecture around it, and a governance and security risk assessment run by our own GRC and cyber security teams.
The output is a document you can take to a board. Current readiness, immediate risks and blockers, the process as it works today, where AI and automation could intervene, an option set with indicative costs and licence implications, and our recommendation.
-
Build. Prove it with something real.
Phase two delivers a constrained pilot with success measures, evaluation criteria, and guardrails agreed before the build begins. Not a demonstration. A working proof of value, validated with the people who will use it, plus a scale plan and a benefits model.
This is the step that answers the Department for Business and Trade problem. If you agree the baseline first, you can prove what changed.
-
Security and governance. Inside both phases, not after them.
Security and governance sit inside both phases rather than arriving after them. The operating model, roles, and decision forums. Standards for data and for LLMOps. Preparation for scale.
Where you need full platform readiness before adopting across the estate, standalone security and governance engagements can be added to the work package. That is a choice you make, not a fixed part of the engagement.
The decision gate
This is the single most important thing Surveyor does, and the thing most stalled pilots skipped.
Each phase includes a gate where the technology direction, the delivery model, and the specialists involved are all confirmed. Microsoft 365 Copilot, Copilot Studio, low-code, pro-code, or machine learning. The decision is made with evidence in front of you, before significant commitment.
One team, not a queue of suppliers
Deploying a secure AI agent is not an AI task. It takes six capabilities working together: Microsoft AI Workforce, AI App Development, Power Platform, Cyber Security, GRC, and Infrastructure. Surveyor brings all six as a single team with one point of accountability.
One team
One point of accountability
- Microsoft AI Workforce
- AI App Development
- Power Platform
- Cyber Security
- GRC
- Infrastructure
What you get
- Readiness assessment
- Use case shortlist and prioritisation
- Pilot artefacts
- Governance pack
- Scale roadmap
Who should be in the room
Surveyor sessions work best with the people who own the outcome and the risk. In public sector terms that usually means:
- Chief Digital and Information Officer, or Head of IT
- Chief Clinical Information Officer, in healthcare
- Director of Transformation, or the service director who owns the process
- Senior Information Risk Owner and Data Protection Officer
- Information governance or GRC lead
- The team leader whose people will use the thing
Chapter 07
Where to start
Three recommendations, drawn from what the evidence above shows about the organisations that got this right.
-
Step 1: Choose the process, not the product
Start with a process that has a named owner, a measurable volume, and a person who is fed up with it. Then work out which technology fits. Reversing that order is the most common and most expensive mistake in the evidence.
-
Step 2: Agree how you will measure it before you build it
The Department for Business and Trade could not prove its Copilot pilot worked, partly because it had not set out to measure productivity. The Department for Work and Pensions could, because it built in a comparison group. Decide your baseline first. It costs nothing and it changes everything about what happens next.
-
Step 3: Design governance and security in at the start
Only 30% of government bodies had assurance processes covering AI risks when the National Audit Office looked. Governance is not what slows adoption down. It is what allows adoption to scale past the pilot.
Questions to ask before your next AI pilot
If you cannot answer these, the pilot is not ready.
-
01. Which specific process are we changing, and who owns it today?
-
02. What does that process cost us now, in hours, in money, or in outcomes for the people we serve?
-
03. How will we know if this worked, and what are we comparing it against?
-
04. Is the data this depends on accurate, accessible, and lawful to use this way?
-
05. Who signs off the risk, and have they been in the room yet?
-
06. What happens to the time we save, and who decides?
-
07. If it works, what stops us scaling it next month?
-
08. If it fails, will we know, and will we say so?
Closing
The public sector has already adopted AI. The staff did it before the strategies caught up. What remains is the harder part: turning tools people quite like into services that measurably work better, safely, at scale, and with evidence a board and an auditor will accept.
That is a direction problem, not a technology problem. It is also a solvable one.
FAQ
Why do so many public sector AI pilots fail?
Rarely because of the technology. The published evidence points to use cases chosen after the tool, data foundations that were not ready, governance added too late, and no agreed baseline to measure against. Gartner reports at least half of generative AI projects were abandoned after proof of concept by the end of 2025.
How much time can AI realistically save public sector staff?
Published UK government trials range from 26 minutes a day to no measurable gain, using the same product. The Alan Turing Institute estimates 41% of public sector working time sits on activities generative AI could support. The range is explained by approach, not by the software.
What is AI Surveyor?
A fixed-price Phoenix engagement that takes a public sector organisation from AI ambition to a working, governed pilot. It runs an Advise phase to establish readiness and choose the right approach, then a Build phase to prove value, with security and governance underneath both.
Do we need our data sorted before we start?
No, but you need to know the state of it. The Advise phase includes a data readiness assessment precisely because most organisations do not know. Finding a blocker in week two is far cheaper than finding it in month six.
Source register
Every figure used in this report, with its origin, its publication date, and a link. Nothing is used in this document that is not in this table.
-
41% of public sector working time could be supported by generative AI. 49% education, 33% healthcare, 47% non-frontline, 38% frontline
-
Probation Service 45%, HM Courts and Tribunals Service 44%, HM Prison Service 43%
-
Only 4% of teachers report using generative AI for marking
-
26 minutes a day saved. 20,000 staff, 12 organisations. 13 days a year. 82% would not go back. Over 70% spent less time searching and on mundane tasks. Drafting 24 minutes, presentations 19 minutes
-
19 minutes a day saved. 3,549 licences. 1,716 users compared against 2,535 non-users. Perceived work quality up 0.49 and job satisfaction up 0.56 on seven-point scales
-
No robust evidence that time savings led to improved productivity. 1,000 licences. 72% satisfied or very satisfied. 1.14 Copilot actions per user per day across 63 working days. Slides around seven minutes faster but needing correction. Excel slower and less accurate. 22% identified hallucinations
-
£45 billion a year in unrealised savings, 4 to 7% of public sector spend. 28% of central government systems legacy, up from 26%. NHS trusts 10 to 50%. Police forces 10 to 70%. 15% cannot quantify their legacy estate. Legacy maintenance often three to four times the cost of modern equivalents
-
The £45 billion figure described as worryingly optimistic, and as a reasonable guess that is hard to measure
-
21% of government bodies had an AI strategy, 61% planned one. 30% had risk and quality assurance covering AI risks, 46% planned to. 87 of 89 bodies responded. Legacy systems and data quality identified as fundamental
-
At least 50% of generative AI projects abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value. Earlier forecast was at least 30%
-
Around 95% of enterprise generative AI pilots produced no measurable impact on the bottom line. Based on 150 interviews and analysis of around 300 deployments
-
95% of responding councils using or exploring AI. 7% considered leaders. 71% not ready on workforce skills, 68% not ready on data availability, quality, and storage. Top three barriers: funding, staff skills, staff capacity
-
79% of charities use AI tools in day to day work. 38% actively or strategically, up from 25%. 60% have or are developing an AI policy. 49% held back by accuracy concerns, 49% by data privacy and security
-
208 councils assessed for AI readiness across all four UK nations. Wide variation driven by leadership priorities, legacy systems, and pace of guidance