Beyond the keynote slides: the insurance workflows where AI is quietly producing value today, the ones where it isn’t, and how to tell the difference before you budget.
The insurance industry has now spent a full decade announcing its AI transformation, and consultancies keep the ambition well fed — McKinsey's recent estimates put the value at stake from generative AI alone in the tens of billions of dollars. Underneath the projections, though, the actual deployment record has a clear shape: AI works brilliantly in insurance where the task is narrow, the data is abundant, and a human owns the edge cases. It disappoints where any of those three is missing.
Where it is producing value now
Document intelligence in claims. The single most repeatable win in the industry. Claims arrive as unstructured mess — photos, PDFs, medical reports, police sketches — and models that extract, classify and route this material cut days from cycle times without ever making a coverage decision. The pattern succeeds because the model's job is preparation, not judgment.
Underwriting triage, not underwriting. In commercial lines, AI that reads a submission, pulls the relevant exposures and ranks the queue lets underwriters spend their day on the risks worth their attention. The models that work are decision-support: they compress hours of reading into minutes. Fully automated underwriting of complex risk remains where it has always been — in the demo.
Fraud scoring. Networks of claims, providers, devices and payment routes are exactly the graph-shaped data machine learning likes. Deployed well, models surface anomalies for investigators rather than auto-declining anyone — which is also what keeps them on the right side of fairness obligations.
Service deflection with guardrails. Policy questions ("am I covered for windscreen damage?") are answerable from documents a model can read. The deployments that survive contact with customers keep a hard boundary: information yes, claims decisions never.
Where it still just demos well
- Touchless complex claims. Total-loss motor at the photo-estimation end has made real progress; injury, liability and large property claims have not. Adjusters are not a legacy cost here — they are the product.
- Chatbot sales of advice-heavy products. Models are excellent at explaining products and terrible at owning the consequences of a mis-sale. Regulated advice keeps humans accountable for good reason.
- Pricing autonomy. Actuarial teams use ML features widely, but filed pricing with regulatory accountability is not being handed to a black box in any supervised market we can find — nor should it be.
The test that separates the two lists
Ask three questions of any proposed AI use case. Is the task narrow enough to measure? Underwriting triage has a queue-time metric; "transform claims" does not. Is there enough labelled history? Fraud has decades of adjudicated outcomes; a brand-new embedded category does not — yet. And who owns the error? If the honest answer is "nobody," the use case is not ready for production, whatever the demo looked like.
What this means for embedded journeys
Embedded distribution is quietly one of the best data environments AI in insurance has ever had: structured transactions, verified identities, exact product context, instant feedback on every offer shown. That is why the near-term AI wins in our world are unglamorous and compounding — smarter offer ranking, cleaner pre-fill, anomaly detection on distribution fraud — rather than a robot claims adjuster. The pattern from the wider industry holds at checkout too: narrow tasks, rich data, human-owned edges. Fund those, and the transformation slides eventually become true.