The 5th Hurdle Is Not About AI

The Lloyd's Claims Management Hurdle came into force on the first of January, and the first thing nearly every commentator wrote about was the artificial intelligence. This was understandable. AI is the noise of the moment, and a sign-off regime that asks managing agents to demonstrate human oversight of automated decision-making is exactly the sort of thing the trade press is built to summarise. There were thoughtful pieces. There were fewer thoughtful pieces. There were vendor blogs, of which the less said, the better.

I have spent the past seven months speaking with the people who actually have to put their name to the thing. Heads of claims. Operations leads. Senior counsel. A surprising number of platform vendors. The pattern that emerges has very little to do with artificial intelligence and quite a lot to do with the unglamorous infrastructure that sits beneath it. When I press on what specifically would cause a firm to fail the test, the answer is rarely "we got the model wrong." It is some variant of "we cannot show how the claim was handled across the parties who handled it."

A senior figure at one of the great London Market platforms put this to me with the cheerful resignation of someone who has been saying so for years. "Speed of response is a direct derivative of information held in various units, not integrated. Human effort is required to pull from multiple systems. The data sometimes doesn't match. There are gaps. There's distortion. And it leads to that thing we don't like: interpretation. Is it fact, or is it your view?"

That is the actual content of the Hurdle, read slowly. It requires demonstrable outcomes across six sub-principles, and those sub-principles cohere around something most regimes of this kind in financial services have not previously had to grapple with: outcomes for which a single firm — the lead — is accountable, but which depend on inputs from a chain of others. The broker who placed the risk. The TPA who triaged the loss and the expert who valued the damage. The reinsurer eventually asked how the figure was arrived at with the Lloyd's rules are deliberately clear: the lead agrees the claim, and the followers stay out of it. What the rules are silent on is how the lead is meant to evidence the inputs it relied on, after the fact, in a form a regulator finds satisfying, when each input came from a different firm running a different system.

There is a phrase in the trade for the dollar cost of this gap: seepage. The slow, distributed loss of value occurs as information fails to travel cleanly between the parties who need it. Another senior practitioner described the same phenomenon in a different language. "Frictional costs," he said, and then, after a pause that felt slightly editorial, "The medium has changed across four iterations of digitisation. The data structure has not. The electronic messages we exchange today are, in essence, a reimagining of what was a punch card forty years ago." He was being mildly humorous. He was not, however, being inaccurate.

The Hurdle is not asking the lead to prove it oversaw its model. It is asking the lead to prove it relied defensibly on the work of every party whose output fed its decision. The model is one input among many — alongside the broker's placement, the TPA's triage, the loss adjuster's report, and the expert's opinion. The question the regulator is now asking is therefore not "did your AI behave?" but "can you show what the upstream parties produced, and that what you took from them held up?" To which the honest answer, almost everywhere, is: not in any form a regulator would find satisfying.

What is missing, then, is neither a smarter model nor a better dashboard. It is an infrastructure layer that does for acceptance decisions what messaging standards have already done for the data underneath them. Standards bodies have spent four decades laying very good rails—the cargo moves. What does not yet move, in any standardised form, is the seal on the outside of the boxcar. This small portable record proves who did what to the contents, under what authority, with what conclusion, so that the lead can show its work and the next party in the chain can rely on it without starting from scratch. The platforms that most managing agents already run can, in principle, generate such a record. None of them, in practice, generate it in a form the firm next door can verify.

The better news for anyone who has read this far in mild despair is that the missing piece is engineering, not theology. The schemas exist and the cryptography exists. The institutional appetite, at least in interviews, is unmistakable and often expressed as a sigh. The 5th Hurdle is, for those who would rather it were about AI, a regulatory accident of timing. For those willing to read it more carefully, it is the first British sign-off regime to ask, out loud, why this particular boxcar has never had a seal.

About the Author: Alexander Barrett is a Senior Fellow at the Verida Charter Foundation, where he researches multi-party ecosystems — how verification work travels across firm, platform and industry boundaries in insurance, supply chains and customs. Over the past year he has spoken with claims practitioners, operations leads, platform vendors and standards-body executives across the London Market and beyond. He was previously a research fellow at Georgetown University and has spent nearly a decade building coordination technology for global supply chains. He writes in a personal capacity.