Government organisations risk overlooking effective, established technologies in their rush to adopt the latest generative AI models head of data science UK&I at Atos
Speaking during a discussion at Think AI for Government 2026, Perrott (pictured) argued that traditional machine learning, statistical analysis and optimisation techniques remain valuable tools for improving public services, often offering advantages in cost, performance and transparency.
“Consider everything in your toolbox, and don’t throw out some things just because they’re a bit older,” he said.
Perrott, who is working with the Department for Education (DfE) on projects ranging from statistical modelling to agentic AI, said public bodies should select technologies according to the problems they are trying to solve.
Choosing the right technology
While acknowledging his own enthusiasm for emerging technologies, Perrott cautioned against assuming that large language models were the best option for every application.
For forecasting, established machine learning techniques may provide better results, he said. Statistical methods remain useful for drawing conclusions from data, while traditional optimisation techniques can offer cost-effective ways to improve processes. These approaches can also be easier to evaluate and less susceptible to problems such as hallucinations.
“Take a very broad approach,” he said, encouraging departments to consider the full range of available technologies.
Generative AI has particular strengths when dealing with unstructured information, such as lengthy documents written in natural language, or when outputs need to be accessible to non-technical users.
However, when a system only needs to produce a numerical forecast for another system to act upon, Perrott suggested that a conventional model with measurable performance might be more appropriate.
Building systems that can change
Perrott also urged government organisations to distinguish between the technology foundations that need to last and the AI tools that may quickly become obsolete. Data structures, infrastructure and governance arrangements require long-term investment because they underpin multiple projects and future developments.
Perrott suggested that organisations should design systems so they can remove one model and introduce another without rebuilding everything around it.
He also warned that spending too much time perfecting an AI component could be counterproductive if the technology is likely to be replaced within a relatively short period. In some circumstances, deploying a useful system several months earlier could deliver greater value than investing in refinements that would soon become redundant.
Measuring outcomes, not just efficiency
Perrott argued that AI projects should begin with a clear understanding of the outcomes organisations want to achieve.
“I would suggest that the best way to begin is if you start with this is what we want to achieve, and work backwards,” he said.
That means identifying the desired improvement, determining the decisions and actions needed to deliver it, and then establishing the data and technology required.
He suggested that departments should assess projects against consistent measures, allowing them to compare performance and make informed investment decisions.
Crucially, those measures should extend beyond the time and money saved by introducing automation.
“Every government department, every piece of government is providing something for the population of this country, and that’s its final purpose,” he said.
Transparency and consent
Elsewhere, Perrott acknowledged that scepticism about AI was not entirely unfounded. He argued that organisations should be open from the beginning about why they are deploying AI, who stands to benefit and how success will be measured.
That transparency would help people understand the choices being made and allow external scrutiny of whether systems are performing as intended. He also suggested that ethical AI was not always straightforward, particularly when different measures of fairness could produce conflicting conclusions.
Rather than presenting such decisions as purely technical, organisations should explain the trade-offs involved.
Asked how government could improve public perceptions of AI, Perrott identified two priorities.
“Success and consent would be my two key things there,” he said.
He warned that people could resent AI when they felt it was being imposed on them without any meaningful choice.
“I think people don’t like having AI being foisted on them,” he said.
Offering AI-enabled services as an optional alternative, and ensuring they genuinely improve the user experience, could help build confidence.
Ultimately, he argued, the case for AI in government would be strengthened by demonstrating improvements that citizens could recognise. Departments should therefore judge their investments not by how advanced the technology appears, but by whether it delivers better services for the people who depend on them.
