In Depth
The defining question of any liability policy is which event triggers coverage. A claims-made policy answers "the claim." Coverage attaches when a claim is first made against you and reported to the insurer, both during the active policy period, not when the act that caused the harm happened. That decoupling is the whole point, and it's where buyers most often get tripped up.
Two clauses control how far back the coverage reaches. The retroactive date sets the earliest incident date the policy will respond to; anything before it is excluded no matter when the claim arrives. The reporting period (and any tail, or extended reporting period, you buy at expiration) governs how long after the policy ends you can still report a claim. Together they define a window, and a claim only lands if both ends line up: the incident after the retroactive date, the claim during the active or tail period.
This makes continuity the thing that protects you. As long as you keep a claims-made policy in force, each renewal extends the reporting window forward while preserving the retroactive date behind it. Drop the policy without buying tail coverage, and a claim that surfaces the next month about last year's work has nowhere to go. The incident was covered when it happened, but the policy that would have responded no longer exists.
What It Looks Like
Say an AI vendor buys a claims-made policy on January 1, 2026, with a retroactive date of the same day (illustrative). In March 2026 their agent gives a customer a harmful answer. The customer doesn't sue until February 2027. If the vendor renewed the policy and it's still in force in February 2027, the claim is covered: it was made during an active period and the incident was after the retroactive date.
Now change one fact. The vendor cancels the policy on December 31, 2026, and buys no tail. The February 2027 claim arrives with no active policy to receive it. The March 2026 incident was real and once-coverable, but the trigger is the claim, and there's no live policy when it's made.
Why It Matters For AI Vendors
AI claims have a long, quiet tail. A hallucinated output or a subtle bias can sit in a record for months before anyone connects it to harm and files. Under a claims-made structure, that delay is fine, but only if you maintain continuous coverage or buy tail when you switch carriers or wind down. The most common way AI vendors get burned is not a denied claim. It's a coverage gap they created themselves by letting a policy lapse between funding rounds, an acquisition, or a product sunset.