Agentic AI is the game changer. Very few running it yet.

October, 21, 2026 - Baden-Baden, Germany

Walk the Baden-Baden Kongresshaus during this weeks Reinsurance Conference and you hear the same sentence in every meeting: AI will change this industry. The Marsh Re symposium that opened the week put the honest question on its own agenda - has AI developed to the point where all members of the value chain can actually use it, and what trust issues are still holding them back? AXA XL Re’s Bertrand Romagne gave the answer most reinsurers give: useful when applied responsibly, as a support and not a substitution for the judgement of underwriting experts. Jefferies, meanwhile, warned investors that AI adoption in the primary market is quietly eroding demand for reinsurance.

AI is there. The implementation is not.

BaFin’s 2026 risk report finds generative AI in insurance “still in its infancy”, used mainly as chatbots and rarely in pricing or risk assessment. The GDV’s dark-processing rate - the share of P&C processes handled fully automatically - stood at 33.5 percent in 2023, after twenty years of “industrialisation” programmes. Agentic AI, meaning agents that carry a file from first notice to settlement rather than summarising a document on an old screen, is barely out of the pilot phase anywhere in Europe. There is less than a dozen of carriers with agents in full production. None with agents running a whole line.

That gap matters and it has a date on it. Last Thursday, 6,200 kilometres west of the Kongresshaus, Lemonade opened the waitlist for “Lemonade Agentic”: a layer of its operating system built not for customers but for their personal AI agents, which will quote, bind, renew, adjust cover and file claims directly in Lemonade’s systems. Agent to agent, no website, no form, no human. Its CEO had told analysts weeks earlier that machine-to-machine purchase was “very early innings” and that Lemonade was “fine with that”. While Baden-Baden debates whether AI can be trusted inside the value chain, one carrier has decided the customer’s side of the counter will be an agent too. The debate about trust has to prevail, but not at the expense of speed.

This gap exists because the industry is following the same playbook it used during its first industrialisation

In the 2000s, insurers standardised what could be written down as rules, automated it, and routed everything else to human specialists through a workflow tool. It worked - up to a third of the factory. The other two thirds needed judgement, and software could not hold judgement. Today the same reflex produces a copilot in the claims system, measured in minutes saved per handler. That is an assistant inside an unchanged process. It will plateau exactly where dark processing did.

Agents are different in kind. An agent receives a goal and a mandate, reads the case, decides, acts, checks and decides again. It reaches the part of the factory the first revolution never touched. But only if the process is rebuilt for it - and that is organisational work, not a procurement decision. Which is why it is early, and why it is slow.

On Monday evening in Berlin, 620 km northeast of Base-Baden, Chancellor Merz and Jensen Huang announced a joint initiative on “physical AI” for German industry. Huang’s framing is worth citing. The next wave of AI, he said, leaves the screen and operates the factory. Every company will have two factories, a real one and a digital twin. Data and domain expertise are a nation’s natural resource. And: “This industrial revolution is so mighty, so powerful - you don’t have another chance.”

Insurance has no robots. But it has a factory: the file.

Baden-Baden exists to price risk at least twelve months ahead of the loss. Agentic AI deserves the same discipline. The foundation models are rented by everyone - that is beta. Alpha is the carrier that moves the judgement of its best people onto its own data in agents, with mandate, decision log and six-eyes control built in from day one.

The good message for Baden-Baden delegates: The Agentic AI in production is agnostic to cycles. The competitive advantage will be with those that implement fast.