AI doesn't make the old claims assembly line faster. It replaces it. Here's what actually changes — and why retrofitting can't get you there.
Every third-party administrator runs, underneath, the same model. A claim comes in. It's logged, triaged and queued. A handler picks it up, works it through a sequence of manual steps — chase the documents, check the cover, set a reserve, negotiate, settle — and periodically, a file goes back to the carrier summarising what happened. The technology has changed around the edges over thirty years. The model hasn't. It's an assembly line, and its throughput is set by how many trained people you can put on it.
That model is now being invalidated. Not improved — invalidated. And the distinction matters, because most of the industry is making a category error about what AI does to claims.
The instinct, when a powerful new tool arrives, is to point it at the existing process. Add an AI step to the assembly line: an AI that reads the documents, or scores the fraud risk, or drafts the letter. Each of those is real and useful. None of them changes the model. You still have a queue, still have a handler working a sequence, still have a file going back monthly. You've made a station on the line a bit faster. You've not changed the line.
This is why "we've added AI" and "we're AI-native" are not the same sentence. The first is an incumbent bolting a capability onto a model designed before the smartphone. The second is a different model.
Three things change when you build the model around AI instead of bolting AI onto the model.
The work routes itself. In the old model, every claim enters the same queue and waits for a human. In an AI-native model, every claim is assessed at first notice — complexity, value, fraud signal, coverage clarity, vulnerability — and routed. The clear, low-value, low-risk claims run straight through under guard-rails and settle in a single session. The complex, disputed, high-value or sensitive claims go to a human expert who has the time to do them properly, because they're no longer buried under the routine. The handler stops being a processing station and becomes what they should always have been: a decision-maker on the claims that need judgement.
The data becomes live. The monthly bordereau is an artefact of the old model — it exists because the handler's work only surfaces to the carrier when someone compiles it. When the platform runs the claim, the carrier's view of their book is live by default. There's nothing to compile. You move from a spreadsheet that lands weeks late to a data plane your own systems — and your own AI agents — can query in real time.
The intelligence is accountable by construction. This is the part that separates a claims platform from a chatbot with a claims skin. In an AI-native model built properly, the AI can't invent a number: the figures come from executed queries against your real data, not from a language model's imagination, and every answer carries its source. AI actors are permissioned and audited exactly like human users. Nothing reaches a decision without a citation and an audit trail. The AI is powerful precisely because it's constrained — the opposite of the "black box" carriers rightly fear.
You cannot reach this model by adding features to the old one, for the same reason you cannot reach a new building by renovating the old one on the same foundations.
The routing, the live data plane, the accountable-by-construction intelligence — these are architectural decisions. They have to be true from the first line of code: how the claim record is structured, how the data is mirrored, how an AI action is authorised and logged. Retrofit them onto a twenty-year-old core and you get a compromise that satisfies neither the old model nor the new one.
That's not a knock on the incumbents. It's a structural fact, and it's the reason a well-built new entrant can move faster than a well-funded old one. The incumbent's installed base is an asset for selling and a liability for rebuilding. Ours is a blank page — which is a liability for selling and an asset for rebuilding.
None of this stays a secret. Within eighteen months, "AI-native" will be a claim every TPA makes, and buyers will have learned to tell the difference between a model rebuilt around AI and a feature bolted onto a legacy one. The advantage belongs to whoever builds the real thing first, in the market where the conditions are sharpest.
We think that market is UK motor, and we think the moment is now. That's a separate argument — and one we'll make next.
ADJST is building the AI-native TPA for UK motor. We're pre-seed, and looking for the founding partners and investors to build it with us.