Machine learning underwriting is the use of algorithms trained on historical and current data to generate risk predictions, pricing signals, and triage scores that manual review alone can't produce. It's one of the more established applications of predictive analytics in commercial P&C, sitting alongside the GLMs actuaries have long relied on, and it's the subject of the first half of this piece: what it is, how it compares to a GLM, and where it still runs into real limits.

Machine learning underwriting is also one part of something larger. AI underwriting is the broader term for how AI is reshaping underwriting as a whole, and it covers generative AI, which reads and summarizes submissions, guidance documents, and policy wording, and agentic AI, which plans and carries out multi-step underwriting work across systems, alongside machine learning itself. To put the scale of that wider shift in context, PwC's 2023 Emerging Technology Survey found AI was the single most-funded technology priority among the US companies it surveyed, with more than two-thirds already running or planning generative AI work alongside efforts like the predictive modeling covered below. The second half of this piece covers that wider shift, and where it's actually showing up in underwriting today.

What is machine learning in underwriting?

Machine learning underwriting applies models to submissions, claims history, pricing variables, and underwriting notes to generate signals for risk prediction, pricing adequacy, appetite fit, and triage speed. It differs from GLM-based pricing in one way that matters: it learns non-linear relationships between variables directly from the data, rather than requiring actuaries to test and encode each interaction by hand.

How does it differ from the GLMs actuaries already use?

GLMs capture relationships between a response variable and explanatory variables through a link function, which is why actuaries trust them: coefficients are interpretable and the math is auditable. The trade-off is that a GLM only picks up the non-linear effects an actuary has already thought to build as a feature. Machine learning shifts that work to the algorithm itself. Tree-based models and neural networks surface interactions the actuary never specified in advance.

A 2026 study comparing tree-based ensemble methods, including bagging, random forests, and gradient boosting, against a Tweedie GLM benchmark on real motor insurance claims found the ensemble methods delivered modest but measurable improvements in predictive accuracy. The gain is directional, not transformational. GLMs remain a credible baseline, and the case for machine learning underwriting rests on where those modest gains compound across a large book of business.

Why underwriter judgment still matters in commercial P&C

Commercial exposures resist full automation more than personal lines. Business nature, financial stability, claims history, and broker context vary in ways that don't reduce cleanly to structured features. The underwriter, not the model, still makes the bind or decline decision. In commercial P&C, machine learning underwriting works best as decision support: it prepares the evidence, and the underwriter applies broker context, appetite fit, and judgment before quoting.

Where the predictive layer runs into limits

Even a well-built model runs into two problems that don't go away with more validation at launch.

Data quality: inconsistent inputs undermine even a well-built model

Model performance depends entirely on the consistency and completeness of the data feeding it, and underwriting data rarely arrives that way. It sits scattered across submissions, pricing tools, claims systems, spreadsheets, and broker emails, so inconsistent capture or missing fields weaken predictions no matter how sophisticated the model is.

Model drift: yesterday's conditions don't stay accurate

A model trained on past conditions can misprice risk as the portfolio, market, or claims environment changes, and it won't announce that it's become wrong. Without ongoing monitoring, a model keeps pricing off stale relationships long after they've shifted, which is why drift needs a recurring monitoring schedule rather than a one-time validation at launch.

What's actually changing underwriting right now

The louder story in underwriting right now is generative and agentic AI: two things that are different from machine learning, and from each other.

Generative AI reads and produces content, summarizing a submission, extracting data from a loss run, or answering a question against a guidance document, working with unstructured material on request rather than predicting a risk score. Agentic AI goes further: it takes a goal, breaks it into steps, and carries out the work across systems, with a human reviewing the exceptions rather than every case. A 2026 academic study of straight-through underwriting for small commercial policies tested this directly: a multi-agent "agentic RAG" system that combined targeted document retrieval, third-party data checks, and explicit rule evaluation outperformed both a single-LLM baseline and a simpler retrieval setup, with the largest gains on multi-step and missing-information cases, exactly where the simpler systems made unsupported decisions.

The investment is already following this shift. PwC's 2023 Emerging Technology Survey found AI was the single most-funded technology priority for the US companies it surveyed, and more than two-thirds of respondents had already implemented generative AI or planned to within the next 12 months. NAIC's own survey data shows the same pattern inside insurance specifically: a majority of P&C, life, and health insurers now report using, piloting, or planning AI and machine learning across marketing, underwriting, pricing, and claims, with underwriting use concentrated in renewal evaluations and policy verification rather than ground-up model-building.

A concrete example of where this actually lands: Allianz UK built a generative AI tool, nicknamed BRIAN, to let underwriters ask plain-language questions against roughly 70 guidance documents, some running 600 pages, and get a sourced answer with a link back to the original page, instead of searching manually. After a pilot with 190 users and nearly 3,000 questions, it moved into production across 260 underwriters, with a plan to reach Allianz UK's full 600-plus underwriter base. Nothing about BRIAN predicts a risk or sets a price; as the transformation lead behind it put it, "this is not about replacing decision making, but equipping people with better tools to access information they need, at the moment they need it." That's a notably more modest claim than the autonomous framing some of the same industry is reaching for elsewhere: Allianz's own technology blog describes a next stage of "agentic AI" that "doesn't just assist underwriters but actually performs underwriting itself", a direction the academic research above treats as some way from production reality, where humans still review the exceptions rather than the system running unchecked.