Most people's experience of AI so far is a system that answers a question or drafts a document. Agentic AI goes a step further. An agent can plan a task, make decisions along the way, use other software and complete the job with limited human involvement. In most industries that is an interesting productivity story. In financial services and insurance, it is a rather bigger deal. A system that summarises a policy document is one thing. A system that influences an underwriting decision, moves money or makes a recommendation with financial consequences is another.
On 23 October 2025, Walmer Group and Blott brought founders, investors, executives and operators together at the London offices of Dawn Capital to work through what this means in practice.
The afternoon had three parts:
- A fireside discussion with Simon Rowell, Senior Vice President at Nscale, on AI infrastructure, sovereign compute and why the UK is an attractive place to invest.
- A financial services panel with Ari Cohen (Macquarie Group), Angus MacNee (Rimbal and ValidPath), Brian Wagner (Revenir) and Matt Lonsdale (Davies), moderated by Jasmine Baker MBE of London & Partners.
- An insurance panel with Steven Mendel (ManyPets), Sam Clifton (Greenlight Re), Nrithya Saravanabhavan (Aviva) and Scott McKenna (careless insurance), moderated by Maria Josife of Erevena.
What were the main themes?
Infrastructure comes first
The question in the room was no longer "what can AI do?" It was "what do we need in place to scale it safely?" Agents depend on reliable compute, clean data, sound integrations and proper security. Without those, a clever pilot stays a pilot.
Sovereignty and trust
Where does the data sit? Who can see how the model reached its answer? Is there an audit trail a regulator would accept? Several speakers argued that control over data jurisdiction, model transparency and auditability will decide who wins, particularly in regulated markets. One line from the fireside stuck: compute may be the new oil, but sovereignty decides who gets to refine it.
Decision quality beats speed
Automating a slow process is the obvious use. The more valuable one is making better decisions: pricing risk more accurately, spotting fraud earlier, giving a customer the right answer first time. Speed is easy to measure. Decision quality is where the return is.
Humans stay in the loop, in a different seat
Nobody on either panel suggested removing people from decisions involving customers, money or regulatory obligations. What changes is the job. The future workforce manages agents, not spreadsheets. That needs new skills, clear accountability and a plan for what happens when an agent gets it wrong.
Why are financial services and insurance a special case?
Three reasons came up repeatedly.
- Liability. If an agent makes a poor decision, who is responsible? The firm, the vendor, the person who approved the deployment?
- Explainability. Regulators and customers expect to know why a decision was made. "The model said so" is not an answer.
- Control. Firms need to be able to limit what an agent can do, monitor what it did and switch it off.
None of these are a reason to stand still. They are reasons to build governance at the same pace as capability.
Does every firm need to deploy agents now?
No. Deploying agents because a competitor issued a press release is not a strategy. Every firm does need to understand where the technology could change its economics or its customer experience, and to run controlled experiments in those areas. The firms likely to gain an edge are the ones that test what works, measure it honestly and build the controls alongside.
The advantage may not come from the technology at all. It may come from organising the data, the people and the governance around it better than the firm next door. How many boards are asking that question yet?
Frequently asked questions
What is the difference between generative AI and agentic AI?
Generative AI produces content such as text, images or code in response to a prompt. Agentic AI can plan and carry out multi-step tasks, make decisions and interact with other systems with limited human involvement.
How is agentic AI used in insurance?
Early uses include claims triage, fraud detection, underwriting support, customer service and policy administration. Most deployments keep a person responsible for decisions with financial or regulatory consequences.
What are the main risks of agentic AI in financial services?
The main risks relate to accountability for decisions, the ability to explain outcomes, data security, regulatory compliance and operational control. Strong governance, audit trails and human oversight reduce those risks.
Do companies developing agentic AI qualify for R&D tax relief?
They may, where the work seeks an advance in science or technology and involves resolving technological uncertainty. Integrating an off-the-shelf model is unlikely to qualify on its own. Developing new methods, architectures or capabilities may.
Key takeaways
- The conversation has shifted from what AI can produce to what AI can do, and what is needed to scale it safely.
- Control over data location, model transparency and audit trails is becoming a competitive advantage.
- The biggest returns come from better decisions, not only faster processes.
- People will increasingly manage agents instead of spreadsheets. Oversight is a design decision, not an afterthought.
