The Engine That Prices Risk

Articles

By Mark Wasuwanich · September 5, 2026

Insurance is one of the largest industries in the world, and everything downstream of the price is downstream of a single number that a small group of people set once a year, using tools built before the internet existed. That number is almost always wrong in the same direction. We've spent the last few months reading through what people are actually trying to build in and around insurance pricing, and this is what stuck with us.

What's already been tried

The most obvious idea is to point a large tabular foundation model at insurance data and let it run. Someone actually tried this earlier in the year with TabPFN, the leading model in that class, and a plain generalized linear model won three of four motor pricing benchmarks. TabPFN won none of them [1]. That's not proof foundation models are a dead end here, it's proof that a model pretrained on generic synthetic tables doesn't automatically understand insurance risk, and handing it a pricing task with zero adaptation was never a fair test to begin with.

The more interesting work happens when the language model isn't the one setting the price. A Toronto group used an LLM purely as a feature engineer, feeding its embeddings into a classical GLM and letting the GLM do the actual pricing. In thin-data regimes, that cut Poisson deviance by roughly seventy percent versus a plain GLM [2], which is exactly the position a new MGA finds itself in: not enough history, and a real need for a number they can defend. Around the same time, a group at UNSW built something called the Actuarial Neural Additive Model, a network designed from the ground up around what regulators actually ask for: per-feature subnetworks, lattice regression to force monotonicity, a three-stage process for deciding which effects are even allowed to show up. It beat every baseline they tested it against on Belgian motor data, and every effect the model learns can be drawn as a chart and handed to a room full of actuaries [3].

The pricing model itself is shifting too. A multi-agent underwriting pipeline with retrieval and a built-in reflection step scored 86.5% decision accuracy on small commercial policies, against 77.6% for a single LLM working alone, and most of that gap came from the multi-agent system correctly declining to decide when the evidence was too thin [4]. Separately, a research group at NYU has started trying to formalize how you'd even price insurance for AI systems themselves: a five-part risk state for agentic AI deployments, with a full contract structure built on top of it [5]. A new category of insurable exposure is arriving alongside AI deployment, and people are only just starting to build the vocabulary to underwrite it.

Underneath all of this sits a scaling story moving faster than most of insurance has clocked. TabPFN has gone from toy tabular problems to million-row datasets across its 2.5 and 3 releases, picking up test-time compute scaling and support for relational data along the way [6, 7]. And a 2026 paper on scaling laws in actuarial ratemaking found that a far simpler architecture, TabM, scales with more data noticeably better than a tuned transformer does [8, 9], which cuts against the easy assumption that whatever eventually wins here has to look like an LLM. The architecture question is still wide open.

What's been learned

Put it together and we don't think the opportunity is a better ANAM. We believe it's an insurance-native foundation model, one whose internal representation actually learns the structure of insurance risk, with ANAM-shaped components sitting on top as one of several constrained output heads, rather than the whole architecture.

The distinction matters because risk and price aren't the same object. Price depends on expected loss, expenses, capital, reinsurance, margin, regulation, appetite, geography, distribution channel: a long list of business decisions stacked on top of risk. Risk is the thing underneath all of it. A model that tries to output a premium directly is quietly conflating all of that into one number, and it inherits every pricing decision baked into whatever historical data trained it. A model that instead learns a shared representation of risk, with separate heads for premium, underwriting decisions, reserves, renewal probability, and tail estimates, keeps those things where they actually belong.

Pretraining is where this gets interesting. Insurance data isn't a flat table. It's relational and temporal: customers, policies, coverages, claims, payments, renewals, and endorsements, all linked to each other across time. And underneath the surface differences between lines, the same math keeps showing up: frequency, severity, duration, selection, dependence, tail behavior. Motor, workers' comp, homeowners, life: different features, same underlying problems. That's exactly the kind of repeating structure a pretrained model can learn, if the training data is built for it. The safer path is probably not pretraining directly on historical insurance data, which risks just memorizing how ten insurers happened to price their books. It's building a generative prior over synthetic insurance worlds, in the spirit of how TabPFN itself was originally trained, and adapting that to real portfolios afterward.

An architecture that could actually earn regulator sign-off probably has to layer three kinds of data into one shared risk state. Structured policy and claims data goes through a tabular encoder. Text, inspection reports, claim narratives, goes through a language encoder. Event sequences, renewals, claim payments over time, go through a temporal encoder. All three feed a shared latent representation of risk, and the constrained heads sit on top of that: an ANAM-shaped head for the technical premium, with monotonicity and smoothness enforced mathematically; a distributional head for tail loss and calibrated intervals; a survival head for renewal and lapse. Each head keeps the interpretability regulators want, while still drawing on a much deeper shared representation underneath it.

None of these pieces are speculative. Every one of them showed up in a peer-reviewed venue or a serious preprint in the last twelve months. What's missing is the integration, the pretraining strategy that stitches it all together, and a book of business to actually run it on.

Looking forward

We believe the pricing engine that reshapes commercial insurance over the next ten years looks something close to the stack above. It gets built by a small team that treats the actuarial toolkit as a foundation instead of a constraint, ships a first version in some narrow line where data is thin and the interpretability bar is high, and expands once the model has earned some trust. The interesting business shape isn't one model per insurer, it's a base insurance foundation model with carrier-specific adapters on top, so the shared intelligence layer keeps improving across the industry while every insurer keeps its own data private.

We believe the first real book of business written on something like this shows up before the decade is out, and the incumbents move to acquire, partner, or copy it the moment the numbers start speaking for themselves.

That's the frontier we're watching from Wanich Risk Lab.

References

Wanich Risk Lab, 2026.