Public sector organizations are gaining confidence in adopting artificial intelligence (AI), and more pilot projects are starting to move into production. But despite growing interest in increasingly sophisticated use cases, a lack of trust in AI systems remains the biggest barrier to widespread adoption, according to Kainos’ Head of Responsible AI.
“We are seeing a growing interest in agent AI, for example agent casework solutions and digital twins to support better policymaking,” said Theresa Yurkevich Hoffmann. “Most back-office operations are based on policies and procedures and lend themselves well to orchestration by agents.”
Hoffman said the organization continues to invest in technology that supports document screening, correspondence processing, chatbots and summarization, along with agent AI.
“We are also seeing interest in broader AI work to support sifting, communications, chatbots and summarization. As the market becomes accustomed to this, we expect to see more advanced AI use cases that transform public sector service delivery.”
Much of this work is still in the pilot or minimum viable product (MVP) stage, but adoption is starting to accelerate, Hoffman said.
“While much of our AI work is still in the pilot and MVP stages, we are seeing more projects move into live deployment as customers become more comfortable.”
building trust
Despite these advances, Hoffman believes the biggest barrier to widespread adoption is not the technology itself.
“The biggest barrier is a lack of trust in technology,” she says. “Many pilot programs are successful in a controlled environment, but fail to scale because the organization cannot build trust in the outputs, the processes behind them, or the people who will use them.”
He said organizations continue to raise concerns about bias and accuracy in AI-generated results, uncertainty around accountability between humans and AI agents, and lack of governance over risk levels and approval processes.
Beyond technical issues, Hoffman said, the organization also faces cultural challenges.
“We also see some ‘human’ concerns about the lack of AI adoption, such as work not aligning with principles and values, emotional barriers (e.g. fear of being replaced) and behavioral barriers (e.g. not understanding the ‘why’ or too much information to process). These are often overlooked but can be very important.”
Governance from the beginning
According to Hoffman, successful AI programs tend to start with focused, low-risk projects before expanding more broadly.
“One approach that has worked well is to use the ‘lighthouse’ use case. That is, choose scenarios that have high value but low complexity.
“We then incorporate best practices into design and development to embed responsible AI principles from the beginning, turning this into a repeatable blueprint that can be scaled out.”
He also emphasized the importance of structured workshops to identify potential risks before deploying solutions.
“We have also had great success using structured workshops to surface trade-offs and downsides early, allowing us to educate teams, increase transparency in decision-making, and make them defensible to the broader organization.”
For example, organizations may need to balance automation and human control, speed and accuracy, or availability and sustainability.
If you like this content…
“By having these conversations upfront, you can make conscious decisions that build trust.”
Hoffman also argued that governance should be established early in an AI program, rather than being added later.
“Building governance practices early supports organizations in building AI policies, risk classification and identification systems, and supplier AI management. By doing this early, organizations can have more confidence in experimenting with AI, with confidence that any failures will be detected and mitigated in real time.”
Misconceptions about sovereign AI
Hoffman also believes that organizations underestimate the complexity of delivering AI at scale.
“Many organizations assume that they can extend their existing practices to AI. However, AI brings new challenges that require a dedicated approach.”
He cited areas such as human oversight, evolving regulation, model selection, and sovereignty as examples where traditional governance approaches are no longer sufficient.
He also cautioned that successful AI implementation requires collaboration across legal, data, security, operations, and product teams from the beginning.
“These stakeholders need to be involved from the beginning, otherwise the effort will be fragmented and accountability will be difficult to identify.”
Sovereign AI is another area where organizations frequently misunderstand the challenges, according to Hoffman.
“Many organizations think this is just about where the data is hosted. In reality, it’s about control, assurance, supply chain, and the broader ecosystem of AI.”
He said organizations need to decide what needs to be managed in-house and what can be safely outsourced, such as data, AI models, technology components and skills.
From technology supplier to AI partner
The shift to production AI is also changing what public sector organizations expect from their technology suppliers, Hoffman said.
“Customers are looking not just for technology providers, but for AI partners who can help them operationalize AI.”
Rather than focusing solely on implementation, customers increasingly want support across the AI lifecycle, including identifying and prioritizing use cases, designing solutions, developing testing frameworks, incorporating responsible AI principles, and accelerating adoption in their organizations.
“Organizations not only want to outsource delivery but also build internal trust, there is an increased emphasis on capability building. This is driving demand for more accelerators, playbooks, and repeatable frameworks.”
Hoffman said as sovereign AI becomes a bigger consideration across government, organizations are also seeking advice on resilience, control and national capabilities.
“This reinforces the need for reliable long-term delivery partnerships rather than quick-trade projects.”
