Making sense of agentic AI: real honest insight from insurance leaders

In July 2025, NashTech and Insurtech UK’s Product & Tech group brought together over 30 senior leaders from across the insurance industry to talk about agentic AI. Not another vendor pitch session or framework presentation, but an honest conversation about what’s actually happening when you try to implement these technologies in the real world.  

Strip away the hype, the headlines and the endless promise of "the future", and a more important question emerges: how do we turn all this agentic AI curiosity into real, operational capability? 

So, what is agentic AI anyway? 

Chris Weston, Senior Technology Consultant at NashTech, opened with a helpful clarification: 

"Agentic AI is about building systems that can revise their own instructions, make decisions, and complete a goal — using multiple tools and external sources." 

This isn’t just another chatbot. It’s more like a co-pilot: adaptive, creative, and (mostly) independent. 

Of course, with great autonomy comes… well, a need for serious governance. One attendee summed it up perfectly: “Give it the power to send emails and you better hope it’s not also feeling creative that day.” 

Big potential, bigger blockers 

Simon Clayden, a Principal Business Architect, highlighted that some of the most fundamental customer processes are still surprisingly manual and time-consuming. "Just personally, recently I transferred my pension, which took four months... I don’t understand why it took that long and how many forms I had to fill in."

He explained how agentic AI could help transform this kind of engagement by identifying customer life events, offering tailored support, pre-filling forms, or booking appointments with an adviser. "These things are typically paper-based and take a long time. Could we use agentic AI to understand changing customer needs and proactively offer tailored advice or initiate a process?"

He also discussed another high-potential use case: improving IT operations through autonomous incident management.

We’re thinking about how to move from reactive to proactive incident management. We’ve used monitoring tools, but now it’s about getting agents involved in prediction and remediation  — ideally stopping problems before they happen.

But Simon stressed that the challenges weren’t all technical. 

"We have data and knowledge documents all over the place... they’re not logged in one place. The AI can’t learn if it doesn’t know where to go."

He summed it up plainly: 

"It’s not a tech problem. It’s everything that goes around it — policies, explainability, governance, and getting the right people involved who understand the process."

Others agreed. One leader shared how their simple document indexing pilot stalled because internal security told them to run it on their own laptop first — with unclassified documents. 

So while the tech is ready, the organisation often isn’t. 

What’s really getting in the way? 

Pete King from OIP InsurTech pointed out a common blocker:

“CEOs love talking about AI. But when you ask a couple of questions, you realise it’s often about education.” 

A few themes surfaced again and again: 

• Data sprawl – knowledge is scattered, siloed, and hard for AI to learn from 

• Lack of clear governance – who’s accountable when AI gets it wrong? 

• Talent gaps – not just engineers, but prompt engineers and hybrid thinkers 

• Overloaded internal teams – compliance, legal, and infosec teams are maxed out 

• Leadership pressure – boards want guarantees, explainability, and fast ROI 

Coralyn Co, Head of Finance Transformation at Hastings Direct, shared how she’s building real traction with agentic AI — not by treating it as a side project, but by embedding it in live transformation work and setting up a dedicated think tank: 

“We’ve started to hire people that would have development of agentic AI in their job description... We want to form a SWAT team around this as we transition our technology architecture - not a big team with lots of bureaucracy, but instead high calibre, highly focused, and resourced enough that it's part of their main role.”

She also highlighted the challenge of securing executive approval, particularly around making sure the estate stays safe and auditable as autonomy increases. 

From side hustle to business strategy 

One thing became clear: success isn’t coming from flashy, standalone AI projects. It’s coming from embedding agentic AI into broader transformation programmes — finance, operations, customer journeys — as an enabler, not the headline act. 

David Mitchell-Dawson, Associate Director of Product & Innovations from EIP, shared some of the unique challenges smaller companies face when exploring agentic AI: “It's not that we don't know what the use case is — it's how do we get our customers to trust that? Because ultimately, they'll be using it. 

He spoke about the dilemma of whether to build or buy: 

“Buy is tough, because the big players are geared for large enterprises. But the cheaper alternatives often haven’t really pulled it through either.” 

His team is already using tools like Claude and Lovable to boost internal productivity: 

“Our developers are using Claude — it massively increases productivity. And we’re using Lovable for prototyping — it really speeds us up.” 

But productising those capabilities for customers remains a work in progress: 

“That’s probably our biggest barrier. How do we turn these internal gains into something usable by our customers? That’s where we still have a lot of work to do.” 

The challenge isn’t knowing what the use case is — it’s building something your end-users actually trust and want to use. That’s where the real work begins. 

It’s not just tech — it’s culture 

Coralyn, again, brought the group back to reality. Even with momentum, progress can be glacial: “Months later, we’re still trying to move forward. It’s not about desire, it’s about time, clarity and the right team around the table.” 

Her solution? Integrate agentic AI into existing transformation work, don’t treat it like a side project. 

Jeff Mountford from Markerstudy Distribution referenced a colleague’s comments: “Instead of looking at it as an AI technology project, we view it as business process transformation. If we add AI into our existing process, we’ll get a small uplift — but to really transform, we need to rebuild the process with AI at the heart of it.” 

He shared how framing the work this way helps gain buy-in from business stakeholders: 

“Forget AI — it could be anything. We’re just trying to fix a process. When the business sees it that way, they’re far more on board.” 

Jeff also noted the organisational reality they face: 

“People look at our tech landscape — with different policy systems, digital estates, data sets — and feel overwhelmed. So, we try to flip the narrative and focus on transforming one process, on one system, on one data set at a time. That’s where momentum builds.” 

So, what’s actually working? 

A few bright spots emerged, and they shared the same DNA: 

• Start with a clear, real-world business problem 

• Integrate into wider transformation efforts 

• Set realistic expectations about autonomy 

• Secure senior buy-in, especially from legal and compliance 

• Define ownership, roles, and measurable outcomes 

The pattern? These weren’t AI-first projects. They were business-first problems — solved with AI. 

Where do we go from here? 

This wasn’t a one-and-done event. Attendees agreed to continue the conversation, sharing use cases, blockers, and lessons learned.  

The focus now is action. Not more hype. Not another deck of aspirational ideas. But real, operational traction. 

Just because agentic AI can do something, doesn’t mean it should. 

But with the right problems, the right people, and the right frameworks? It has the potential to fundamentally reshape how insurance works. 

 Want to join the next Product & Tech group roundtable or talk through your blockers, ideas or goals? Get in touch here.