AI Utilisation and How to Guard Against Inherent Risks

Insights from a panel moderated by Max Richter, EMEA CEO at mea Platform, with David Griffiths (Citi), Vishal Marria (Quantexa), and Pravina Ladva (Swiss Re). 

Key Takeaways

• Most organisations have moved past the question of whether AI works. Few have moved past pilots into consistent, measurable ROI 
• A core underwriting process at Swiss Re collapsed from 45 to 50 days to one day using 14 AI agents with governance and controls in place 
• AI value depends on data trust; in regulated industries, the gap between plausible and defensible output is the gap between a pilot and a production deployment 
• Change management accounts for roughly 60 percent of whether an AI programme succeeds; technology alone does not drive adoption 
• As AI systems gain autonomy, governance must be continuous and proportional, not a single gate at deployment

Why is most Enterprise AI Still Stuck Between Experimentation and Value?

The volume of AI activity inside large, regulated organisations is no longer the problem. Citi has 190,000 users and 30,000 developers on AI tools across 160 countries. Swiss Re has fundamentally reimagined some of their core insurance processes from end to end.  

The proof points are real, but what remains inconsistent is the return.  

Many organisations now run hundreds of use cases. That type of breadth with insurance AI helps build institutional learning, but it does not automatically produce business value. Treating AI as a business transformation programme rather than a technology initiative is what is separating those who are succeeding at scale, from those finding small, fragmented successes. 

Pravina Ladva put it plainly: experimentation is high, but ROI is not yet consistent. The issue is focus. AI has to move from the innovation lab into core business strategy before the economics follow. 

For (re)insurance, the pattern is familiar. The industry has pilots. What it lacks is pilots that compound into structural improvements in combined ratio, margin, and operating leverage. 

How Much can Agentic AI Actually Compress an Insurance Workflow?

A core underwriting process at Swiss Re that used to take 45 to 50 days now takes one day. As Pravina described it, “Underwriting that used to take 45–50 days now takes one day, using 14 agents working together with strong governance and controls.” Swiss Re has moved from legacy ways of working to AI-native workflows and seen incredible results thus far. 

A 45-day process collapsing to one day is not a productivity improvement. It is a different operating model entirely. The underwriter’s role, the broker relationship, the capacity calculation, the cost per (completed) transaction all change when work stops being sequential and starts being orchestrated by AI agents. 

In the past, workflow automation from insurtechs or in-house builds has historically meant making slow processes slightly less slow. But agentic AI, properly deployed, makes them structurally different. The gains compound with every transaction.

Why Does Data Quality Determine Whether AI Scales or Stalls? 

AI value depends almost entirely on the trustworthiness of the data underneath it. When models are grounded in contextualised, governed data, results improve. When they are not, the output may sound reasonable but will not survive scrutiny from a regulator, an auditor, or a board. 

Vishal Marria was clear on this – garbage in, garbage out is still true; and it lands differently in a regulated industry. The requirement is more than clean data; it is contextualised, governed, transparent data that can defend decisions at scale while operating under the businesses’ same decision-making process. 

For (re)insurance, this is where insurance-native intelligence matters. A general-purpose model fed insurance documents will produce plausible answers. A model trained on insurance workflow logic, anchored in a proprietary knowledge graph, will produce defensible ones, and with higher accuracy. In (re)insurance, that distinction determines whether a deployment stays live or does not move past a pilot if it makes it to one. 

What is the Biggest Risk to AI Adoption Inside Regulated Organisations? 

The thinking from the panel was that the biggest risk to AI adoption inside regulated organisations is not model accuracy, not integration complexity, not regulatory uncertainty. It is change management. 

Pravina put a number on it: roughly 60 percent of success. So what happens? Getting people to actually adopt and stick with the new way of working is not an easy task. People fear job loss, and that fear must be addressed directly. Without sustained momentum and deliberate workforce engagement, people revert to old habits regardless of how well the technology performs, and regardless of how productive it can make them. 

David Griffiths described the same challenge through Citi’s lens: the focus is on tasks, not jobs. Repeatable tasks get automated. People shift to higher-value work that requires more judgement. The aim, as David framed it, is to build the most AI-empowered workforce globally, positioning AI as a capability for people rather than a replacement of them. 

In (re)insurance, the equivalent is freeing underwriters from intake processing to focus on risk selection, freeing claims handlers from logistics to manage loss outcomes, and freeing finance teams from reconciliation to run analysis. Though none of that happens because the technology works, it happens because someone leads the change. 

How Should Governance Keep Pace with AI Autonomy and other Risks?

As AI systems move from recommending actions to executing workflows, governance must scale with autonomy. The systems that will affect combined ratio are the ones that operate with real agency inside live workflows: ingesting submissions, triaging claims, reconciling finance, routing exceptions.  

That level of autonomy requires governance that is proportional, continuous, and built into the system architecture from the start. 

Pravina described Swiss Re’s approach: assess the autonomy and agency of each AI system, then apply controls that match. Governance runs through the full lifecycle of the workflow, not just at deployment. 

David reinforced the point from Citi. AI risk needs to sit explicitly within the enterprise risk taxonomy. Without that classification, controls drift toward over-restriction that kills the value case or under-restriction that creates exposure. Neither is acceptable in a regulated environment. 

Max Richter is EMEA CEO at mea Platform. mea Operations deploys AI agents that execute workflows across underwriting, claims, finance, and broking.