Your best people are drowning in operational drag; AI will help them win

Martin Henley, of mea Platform, makes the case for why AI that understands insurance could free teams from operational drag and help carriers attend to risk.

Property catastrophe rates fell 14.7% at the January renewals, the steepest reduction since 2014, and global reinsurance capital sits at a record $790 billion. In a softening market with abundant capacity, outperformance comes from one place: understanding risk better than the competition.

That takes your best people spending their time on risk, and today they spend it on your operations. The underwriter re-keying submission data, the broker chasing conflicting numbers across an email thread, the claims handler reconstructing a file; this is the industry’s scarcest resource, experienced judgment, consumed by repeatable work. That is drag. 

The reality is that AI can now do that work properly. For the first time in the history of insurance innovation, operations can be a solved problem. Your people can be cut free from the drag that holds them back, but investment in AI has to be targeted toward tech that makes the difference. 

The industry put close to four hundred million dollars into generative AI across 2023 and 2024, and productivity has yet to follow. The difference between spend that moves the margin and spend that adds cost comes down to a single question: does the technology actually understand insurance?

AI that understands insurance

Insurance operations are fundamentally an insurance language problem. Every transaction is data and wordings moving between steps: a submission arrives, gets decomposed, classified, extracted, reconciled, priced, endorsed, renewed, reserved against. At each step, the meaning of the words carries the risk. A model that understands the language of insurance with exacting accuracy can carry a transaction through every one of those steps with reliability and consistency at scale. Insurance decisions need deterministic outcomes. A plausible guess about a total insured value is a wrong price. A plausible guess about a treaty exclusion is an unpriced exposure.

The failures that matter are specific and expensive. A TIV (total insured value) read from the wrong cell of a schedule prices the risk incorrectly from the first moment, and every downstream decision inherits the error. A broker email, application and statement of values disagree with each other, and the conflict gets silently resolved to one wrong number with no flag raised. A value lands in a pricing model with no lineage back to its source document, and when a regulator asks how it was derived, there is no answer.

Preventing those failures takes a model built on the language of insurance itself. It takes knowing that “TIV” and “total insured value” are the same field, that “limit” means one thing on a statement of values and another on a loss run, that a submission arriving as an email thread with nested attachments is one risk that needs reassembling into a single view. It takes validation against insurance rules, reconciliation across every source document and an audit trail on every field.

To get to this point, where AI agents take the burden of the manual operational load, the technology architecture has to follow from the problem. A language model trained on insurance, paired with a knowledge graph that holds the meaning of the business: the field definitions, context, relationships between them and the validation rules that govern what a value can be. At mea, for example, that graph resolves more than 125,000 insurance fields and applies millions of validation rules, and extraction is one step in a 15-stage pipeline in which every value is validated, reconciled and traced before it reaches an underwriter, a broker or a claims handler.

Run the same submission through twice and you get the same answer, with the confidence score and the lineage to prove it. Speed matters and reliability matters. Consistency is what lets a regulated business put AI into production and defend every output it produces.

Building this takes insurance domain experience. Teams who have spent careers inside underwriting, broking and claims workflows know where the errors hide and what a decision actually requires. Encoding that knowledge into the model layer is years of work. It is also the whole game. 

The repeatable and the consequential

Picture the change at desk level. A submission arrives as an email thread with a scanned application, an SOV and two loss runs attached. Today, someone spends hours pulling that apart and re-keying it. With the pipeline in place, what reaches the underwriter is one structured view of the risk, every value validated, reconciled, confidence-scored and traceable to its source. They open a decision, and the deciding starts immediately.

Scale that across a book and the operating model changes. AI owns the repeatable: intake, extraction, validation, reconciliation, mapping into the schemas your systems require. Your people own the consequential: risk selection, pricing judgment, negotiation, client relationships. The underwriter who currently spends 40 per cent of their time gathering and re-keying information gets that time back for the work that shows up in the combined ratio.

The economics make the case plainly. The back office still consumes 12 to 14 cents of every premium dollar, against an underwriting margin of three to five cents, a ratio that has held for 15 years. Cost reduction there is real but bounded. The prize is bigger: every hour of judgment redirected from operations to risk compounds into better selection, sharper pricing and stronger client relationships. Those are the decisions that determine which carriers are still writing sub-90 combined ratios when this cycle turns, and that is the honest answer to the question every board should be asking about AI investment: did the margin move?

Bermuda has been here before

Bermuda has a habit of moving first when the industry changes structurally. After Hurricane Andrew in 1992, capital formed here to write the catastrophe risk from which the rest of the world was retreating. After September 11, a new class of carriers stood up in months. After Katrina, Rita and Wilma, it happened again. When capital markets wanted a direct route into insurance risk, the ILS market grew up here. Each time the pattern held: Bermuda concentrated underwriting expertise, regulatory pragmatism and capital in one place, and acted while larger markets were still forming committees.

The AI transition rewards exactly those traits. It will be won by the markets that combine deep domain expertise, early adoption, and proximity between the people who make decisions and those who build the tools. Bermuda holds all three in unusual concentration. The underwriting talent that priced the unpriceable after Andrew is the same talent pool that can tell an AI builder precisely what a casualty treaty submission needs to become before anyone can act on it. Decisions that take quarters in London or New York take weeks in Hamilton.

Bermuda has spent three decades proving that a small market with concentrated expertise can lead the global industry through structural change. The AI era is the next test, and the market is positioned to pass it the way it passed the others: by understanding the risk better than anyone else and moving first.

Martin Henley is group CEO of mea Platform. To find out more about mea Platform, visit www.meaplatform.com.

To read the full issue of Bermuda Risk Review 2026, click here. 

Did you get value from this story? Sign up to our free daily newsletters and get stories like this sent straight to your inbox.