Trust remains biggest obstacle to scaling AI, says Kainos

Public sector organisations are becoming more confident about deploying artificial intelligence (AI), with more pilot projects beginning to move into live production. However, a lack of trust in AI systems remains the biggest obstacle to scaling adoption, despite growing interest in increasingly sophisticated use cases, according to Kainos‘ head of responsible AI.

“We are seeing an increase of interest in agentic AI – for example, agentic casework solutions or digital twins to support better policy making,” said Theresa Yurkewich Hoffmann. “Most back-office work is based on policies and procedures, which are a good fit for agentic orchestration.”

Alongside agentic AI, Hoffmann said organisations continue to invest in technologies that support document sifting, correspondence handling, chatbots and summarisation.

“We are also seeing an interest in wider AI work that supports sifting, correspondence, chatbots, and summarisation. As the market gets more comfortable with this, we expect to see more advanced AI use cases that transform public sector service delivery.”

While much of this work remains in the pilot or minimum viable product (MVP) stage, Hoffmann said adoption is beginning to accelerate.

“Much AI work is still in the pilot and MVP stage, but we are seeing more projects move into live deployment as customers get more comfortable.”

Building trust

Despite this progress, Hoffmann believes the biggest barrier to wider adoption is not the technology itself.

“The biggest barrier is lack of trust in technology,” she said. “Many pilot programmes succeed in controlled environments, but they fail to scale because organisations can’t build confidence in the outputs, the processes behind them, or the people using them.”

She said organisations continue to raise concerns about bias and accuracy in AI-generated results, uncertainty around accountability between humans and AI agents, and a lack of governance over risk levels and approval processes.

Beyond technical issues, Hoffmann said organisations also face cultural challenges.

“We also see several ‘human’ concerns around a lack of AI adoption – for example because the work doesn’t align with principles or values, there are emotional barriers (for example, fear of replacement) or behavioural barriers (for example, a lack of understanding ‘why’ or too much information to process). These are often overlooked but can be extremely important.”

Governance from the outset

According to Hoffmann, successful AI programmes have tended to start with focused, lower-risk projects before expanding more widely.

“One approach that has worked well is using ‘lighthouse’ use cases – where we select a high-value but low-complexity scenario.

“We then build best practices into the design and development, and we embed responsible AI principles from the start – and turn this into a repeatable blueprint that can be scaled out.”

She also highlighted the importance of structured workshops to identify potential risks before solutions are deployed.

“We have also seen strong results where we use structured workshops to surface trade-offs or harms early. This helps educate teams and build transparency into their decisions so that they are defensible to the wider organisation.”

For example, organisations may need to balance automation with human control, speed with accuracy, or availability with sustainability.

“Having these conversations up front helps make conscious decisions that builds trust.”

Hoffmann also argued that governance should be established early in AI programmes rather than added later.

“Building governance practices early on – supporting the organisation in building an AI policy, a system of risk classification and identification, and supplier AI management. When we have done this early on, it has made the organisation more confident in experimenting with AI as they have more confidence that any failures will be caught and mitigated in real time.”

Misunderstanding sovereign AI

Hoffmann believes organisations also underestimate the complexity of delivering AI at scale.

“Many organisations assume their existing practices can be extended into AI. However, AI introduces new challenges that require purpose-built approaches.”

She pointed to areas including human oversight, evolving regulation, model selection and sovereignty as examples where traditional governance approaches are no longer sufficient.

She also warned that successful AI adoption requires collaboration across legal, data, security, operational and product teams from the outset.

“These stakeholders need to be involved from the start – if not, the initiative can fragment and accountability becomes difficult to pin down.”

Sovereign AI is another area where organisations frequently misunderstand the challenge, according to Hoffmann.

“Many organisations think this is just about where data is hosted. In reality, it is about control – assurance, supply chain, and the broader ecosystem of AI.”

She said organisations need to decide what they need to control themselves, and what they can safely outsource, including data, AI models, technical components and skills.

From technology supplier to AI partner

The shift towards production AI is also changing what public sector organisations expect from technology suppliers, said Hoffmann.

“Customers are looking for AI partners who can help them operationalise AI – not just technology providers.”

Rather than focusing solely on implementation, customers increasingly want support throughout the AI lifecycle, including identifying and prioritising use cases, designing solutions, developing testing frameworks, embedding responsible AI principles and driving organisational adoption.

“There is a growing emphasis on capability building as well as organisations want to develop internal confidence, not just outsource delivery. This is driving demand for more accelerators, playbooks, and repeatable frameworks.”

As sovereign AI becomes a bigger consideration across government, Hoffmann said organisations are also seeking advice on resilience, control and national capability.

“This is reinforcing the need for trusted, long-term delivery partnerships rather than quick transactional projects.”