AI is moving into the very first step of a claim. This week a low-code platform vendor released an AI agent that takes the first notice of loss for auto insurers, and the same week the trade press reported a rise in AI-assisted fraud aimed at that same step. Insurers should treat these as one design problem, not two separate projects.
What happened
On November 13, OutSystems released Agentic First Notice of Loss, a ready-made product for auto insurers. AI agents take the policyholder’s report, gather driver statements and third-party details, review damage photos as they are uploaded, and check coverage against the policy before opening the claim in the insurer’s existing claims system. Through a partnership with Honk, the agent can also send a tow truck while the customer is still reporting the loss. OutSystems says adjusters keep every decision on coverage, fault and payment, and that every action the agents take is logged. (Insurance Journal)
Two days earlier, Risk & Insurance published a feature on AI and insurance fraud. Verisk’s anti-fraud analytics head described growing use of AI to alter claim photos and documents, and a new pattern of “fraudulent agentic activity”: AI agents that buy policies, file claims through voice bots and cancel once paid. Verisk research found that 36% of consumers would consider digitally altering a claim image or document. (Risk & Insurance)
The same week also brought a proposed exclusive agreement between Mitchell and Tractable. It would let condition reports from Tractable’s vehicle-inspection scans produce preliminary repair estimates in Mitchell’s estimating software. (Carrier Management)
Our take
Automated intake makes sense. First notice of loss is high volume, follows set steps and shapes how the customer feels about the whole claim. OutSystems cites research that 83% of U.S. consumers would switch carriers after a poor claims experience. Taking the customer’s report, checking coverage and arranging a tow in one conversation is a clear improvement.
But the step that is easiest to automate is also the easiest to attack. If a photo goes straight from a phone to an AI that opens a claim, a manipulated photo goes the same way. If a voice bot can file a claim, so can someone else’s voice bot. The fraud controls that used to come later, when an adjuster looked at the file, now have to run at intake, at the same speed as the agent.
OutSystems’ human-approval design helps, because adjusters still decide coverage, fault and payment. But an adjuster reviewing a tidy, AI-written summary may not notice that the photo underneath it was altered. Fraud checks have to run on the original evidence, before anything is summarized.
What to build in
- Check evidence at upload. Screen every photo and document for signs of editing before the intake agent uses it, and send failures to SIU instead of to a summary.
- Tell people from bots. Record whether a claim was reported through a human conversation, an app or an automated channel. Unusual patterns, such as a new policy, a quick claim and a cancellation, should raise a flag.
- Keep the original evidence. Store the raw photos and recordings, not just the AI’s extracted data, so investigators can go back to them.
- Log every agent action. Record what the agent saw, what it decided and what it sent to which system. That trail serves fraud investigators and regulators alike.
- Measure both sides. Track time to first contact and customer satisfaction, and also the referral and confirmed-fraud rates on claims taken by agents compared with those taken by people.
Agents at claims intake will soon be common. The insurers that benefit most will be the ones that build fraud detection into intake from the start, not the ones that add it after the fact.