Government departments must fix long-standing problems with fragmented and poor-quality data before artificial intelligence projects can move from experimentation into real-world use, according to SAS.
Nicola Furlong, the company's VP for EMEA public sector, said attention continues to focus on AI models while the quality and management of underlying data receives less attention.
"You can bolt the smartest model in the world onto a mess - all you get is a faster, more confident mess," she told Think Digital Partners.
"I've sat in enough procurement conversations now to know the pattern: it's the model that gets the attention, but the data foundation can be an afterthought, and then six months later someone's asking why the 'AI project' produced a beautifully-formatted wrong answer."
Government faces a particular challenge because data is distributed across legacy systems and organisational boundaries that have developed over decades, Furlong said. "Government departments need to fix the plumbing first. It's less glamorous than an AI project, but it's the bit that actually determines whether any of it will work."
Most projects still stuck in pilot
Public sector AI deployments are beginning to move beyond experimentation, but the number reaching production remains small. 2025 OECD analysis of almost 1,500 AI use cases collected by the European Commission's Public Sector Tech Watch found that 58 percent were still planned, in pilot or in development.
"Starting a project and running it in production are very different things," said Furlong. There are "plenty of proofs of concept that impressed everyone in the room, and a much smaller number that survived contact with real procurement cycles, real security review, and real budget scrutiny the following year."
Governance presents another barrier, particularly where high-level policies have not translated into practical guidance for the people expected to use AI. "The struggle is in the gap between 'we've written a policy' and 'someone in the building actually knows what it means for their day to day'," she said.
"A lot of governments now have AI strategies and governance frameworks on paper, but very few have built the mechanisms to actually measure whether their deployments are delivering public value."
The OECD's 2026 outlook found that 30 of 36 countries surveyed had at least one institution responsible for governing public sector AI, while measuring AI's impact remains an ongoing weakness. Data sovereignty adds another layer, Furlong said - "Who owns this, where does it sit, who can see it under what legal basis - and you've got the two hardest governance conversations happening at once, usually without anyone senior enough in the room to actually decide."
Synthetic data and fraud prevention
Synthetic data has a practical role where organisations need to work with sensitive information, Furlong said. Government holds tax, benefits and health records that cannot be used directly for training AI systems. Synthetic data can provide statistically representative information without exposing individuals, and can create examples of uncommon events such as specific types of fraud.
"There's an urgent need to address the risks of these models falling into the wrong hands," said Furlong.
SAS research published in December found responsible use of AI and analytics, integration with existing systems, and privacy and security among the main barriers reported by UK public sector fraud professionals. Forty percent said their departments already used AI, while nearly all expected to adopt AI or generative AI within two years if they were not already using it.
Furlong said teams working on fraud, waste and abuse are moving fastest with the technology. "Anything touching individual entitlement decisions - benefits eligibility, immigration, anything where a wrong call has a face attached to it - moves much more slowly, and rightly so."
How to measure AI success in government
Furlong said agencies should also measure AI success by look beyond productivity improvements. "You can save a caseworker two hours a week and still lose if the system is not fair," she said. SAS research found 96 percent of UK public sector fraud professionals surveyed believed fraud, waste and error had negatively affected citizen trust.
Measures of successful AI should include whether deployments reduce errors affecting citizens, allow employees to concentrate on work requiring human judgement, and make public services more explainable, she said.
"Productivity is easy to put in a slide. Trust is the actual score to keep."
Looking ahead, Furlong expects the most important uses of AI in government to become less visible. "I think the honest answer is a lot less 'AI does the thing' and a lot more 'AI quietly catches the thing before it's a problem'," she said.
Why this matters for government
The gap between AI pilots and production deployments directly affects anyone in the public sector responsible for digital transformation or service delivery - and your agency will be judged on outcomes, not on tech demonstrations. The practical takeaway: do not start new AI pilots if your department lacks reliable source in the underlying data and an identified owner. Check which existing pilots can survive the full lifecycle - procurement, security review, and budget scrutiny. If your university is hiring for these roles, build the case for investment in data infrastructure now, rather than after a failed project brings scrutiny.
Your membership also unlocks:
