Featured in Modern Insurance Magazine
By Roi Amir, CEO, Sprout.ai
Insurance has automated the parts of claims that are easy to automate: scanning documents, routing files, extracting data. What it hasn’t done is bring AI into the actual decisioning, the judgment calls that determine outcomes. That’s the real story of the last decade of claims technology, and it’s why AI’s biggest opportunity in insurance is still untouched.
For years, the industry has automated movement, not judgement. Not handling the moment that impacts the customer’s experience and the insurer’s loss ratio: is this covered, what is it worth, and can we defend the answer if challenged. The easy parts were automated, leaving the decision itself locked inside a core system, in a person’s head, or an unread policy document. Automation moved the claim but it did not understand the claim. And powerful general purpose AI has muddled the picture more than clarified it.
That’s why the claim itself should be treated as the product. Not the app, not the portal, not the loyalty scheme wrapped around it, but the claim. It is the only moment in the policy lifecycle where a customer finds out whether the thing they paid for actually works. Handled well, it drives more trust and retention than anything else in insurance. Handled badly, it is very difficult to repair.
If that is true, there has to be a hard line around what we let AI do at that moment. This is where generic AI and purpose-built claims AI stop being a technology debate and become an important question. A large language model asked to assess a claim will produce a fluent, confident answer . That is the problem. A plausible answer is not a defensible one, and that gap is where leakage, complaints, and regulatory exposure live. Ask a generic model to adjudicate the same claim twice and there is no guarantee of the same answer. Ask it to show its working clause by clause, and it will generate an explanation that sounds right rather than one grounded in the policy working. For drafting a letter, that is a minor quirk. For deciding whether someone’s claim is covered, it is disqualifying.
That points to the biggest misconception I encounter: that AI in claims is just a faster version of the automation insurers already have, and that any capable model can be pointed at a policy and trusted with the answer. General-purpose models are genuinely strong at extraction: pulling unstructured data from unstructured documents. Decisioning is a different discipline, applying policy terms, exclusions, and regulatory rules to reach a settlement that is consistent, auditable, and identical on repeated runs, and that improves every time a human corrects it. Confusing the two produces a familiar pattern: a pilot that looks brilliant in a demo, then stalls the moment it meets live complexity, because nobody built the confidence scoring, audit trail, or feedback loop it needs. The model is the easy part; the discipline around it is what makes a decision defensible.
That same discipline should change how the industry measures success. We became very good at tracking handle time and cost per claim, both measures of movement. The metrics that will define who wins the next decade are quality metrics: indemnity accuracy, reserve accuracy, consistency across adjusters and geographies, and whether claims intelligence flows back to underwriting to price risk better. A one percentage improvement in indemnity accuracy delivers more financial benefit than the same improvement in loss adjustment expense. Insurers chasing speed alone are optimising the part of the equation that matters the least.
So if I were advising a global insurer’s CEO on where to invest first, it wouldn’t be a customer facing feature or a general purpose model bolted onto the front end. I would start at first notice of loss, with AI purpose built to read policy wording against claim evidence, because coverage clarity there impacts everything downstream: speed, consistency, and how much the customer trusts the process.
The industry did not spend the last decade automating the wrong things. We automated the easy 80% and left the final 20%, the actual decision, for later. Later has arrived. The insurers who treat the claim as the product they are selling, decided by AI built specifically to be trusted with it, will be the ones customers stay with.
Read the article in Modern Insurance Magazine (page 54)
About Roi Amir
Roi Amir is CEO of Sprout.ai, an AI platform helping insurers, MGAs and TPAs automate claims decisioning and improve the speed, consistency and transparency of claims operations.