‘Agentic AI’ wave: Trust, data and readiness in the spotlight

With talk of agentic AI taking the public and private sectors by storm, how can government prepare its data, governance frameworks and digital infrastructure for this brave new world? It was one question under discussion at Think Data for Government, which examined the key opportunities and challenges for public bodies and their suppliers.

So, with the pace of change accelerating, how can senior leaders avoid being swept away by the hype while still preparing their organisations for transformational technology?

Sumitra Varma, deputy director for data engineering, data science and AI at the Ministry of Justice (MoJ) began by urging perspective: this is a journey, and the public sector is still at the prompting stage.

“Everyone’s excited about AI… but agentic AI is taking it another level. We are really talking about agents that can reason and can think,” she said.

She stressed that the excitement must not distract from the basics: “It’s about the data and the foundations that you have below it. How much you can tune that data, keep it ready for the next – which is actually allowing agents to think.”

Currently, most users are still learning “how to prompt… What if you didn’t need to prompt? What if you could just tell any other tool: go and do this”? The core question, she said, becomes: “How much would you trust it?”

That trust depends on data captured properly from the start. Varma described a recent visit to a prison that highlighted the stakes.

“It was overwhelming… the sheer amount of things that those people on the ground had to do,” she explained. Some tasks were digitised, but some were not. “That made me think; that stuff that isn’t on that digital system, that’s data we’re not seeing.” An agentic system built only on what is captured “won’t know the entire data journey”.

The risk, she warned, is building AI on gaps. “Don’t add data on as a bolt-on at the end… We think of data too late. For senior leaders, they need to see data as part of their entire journey, from the very beginning.”

Varma also emphasised governance as a trust builder. The MoJ, she said, benefits from a “mature team” covering everything from data science to ethics. “That builds trust that we are doing things the right way.”

Frameworks matter too. “When we build algorithm tools, we look at algorithmic transparency reporting standards… so a member of the public can actually go in and look at how we’re building these tools.” But she was clear: “Are we there yet? No… There’s still a lot more to do.”

Context, caution and who’s really the ‘user’?

Jim Stamp, head of technology at Made Tech took a pragmatic view of readiness: “Are we ready for it? Not yet. We’re a long way off.” Even in the private sector, experiments are revealing hard lessons.

He described trying to plug an agentic assistant into HR systems. It unintentionally gained access to all the data for the whole organisation. The issue, he said, is that “they imagine who their user is, but don’t think about actually the agent is a user in their own right.”

On data quality, Stamp challenged conventional wisdom: perfection isn’t essential. “AI… is quite good at coping with bad data.” What matters more is context, he said – being clear about what the system is asked to do and the information surrounding it. “People are advising us to stop learning prompt engineering and just do context engineering.”

Data ownership also needs sharpening. Too often, ownership sits with “only data people”. Instead, “attaching the ownership of those data products to the product owners… was key” when he worked on governance in industry.

Trust, risk and transparency in the public sector

Panel chair Gavin Freeguard asked what makes agentic AI deployment different in government, where people often have no choice over services and where stakes are high.

Transparency, Stamp said, “is probably the most important – but it is contested territory: “Large language models, and even more so agentic systems, are fairly opaque, if not completely opaque, in how they work.”

Testing helps, but only up to a point. Traditional automated testing fails with stochastic systems. Guardrails are essential, too.

Stamp pointed to “loading practices over the top… constraints of why they’re doing the thing”. Constitution-style layers can help: one model produces output, another moderates it. “That double call increases cost” but delivers assurance that content “isn’t causing harm”.

Training models to admit uncertainty is also crucial: “Tell me if you can’t… it fundamentally changes how they work… Telling it to be honest is surprisingly impactful.”

Standards and metadata: the hidden enabler

An audience question highlighted the need for AI standardisation and metadata. For Stamp, these are “fundamental”.

By “data contracts”, he meant clarity about “the schema, the policy under which it was captured… the legitimate use that it can be used for”. Rich metadata boosts correctness and accountability.

Varma agreed, noting that “naughty problems” span departments. She cited a modern slavery data project blocked not by a lack of capability but by a lack of discoverable data. Metadata would have shown “what is the dataset that exists in another department that we could probably tap into”.

Agents will need what humans need: context. “If something was missing, as humans, we’d know. Agents wouldn’t – and that’s where they will make it up,” she warned.