Start with the policy trigger. D&O and E&O policies for AI companies are typically written on a claims-made basis, so they respond to claims asserted during the policy period rather than to loss from an occurrence that took place inside it [5]. Many also carry a retroactive date limiting how far back the conduct behind a claim can have occurred [6]. The two dates that decide who pays are the assertion date and the retroactive date; the day someone noticed the model had drifted plays no part in that calculation [18].
Now run a merger through that structure. On an acquisition it is customary for the target to buy extended reporting period coverage, also called runoff or tail, which prolongs the window for reporting claims about earlier conduct, while the acquirer buys a go-forward policy effective at or around closing for conduct after the deal [7]. Tail policies often exclude claims involving any conduct after the cutoff date, and the go-forward policy may decline the same claim because it also involves pre-merger conduct [8]. A claim alleging wrongful acts on both sides of the cutoff has three possible outcomes, including one where neither policy responds [17]. David Halbreich, an insurance recovery partner at Reed Smith [1], says this catches sophisticated deal and in-house counsel who thought procuring both policies was enough [10]. His fix is to review both policies and their exclusions in tandem, then get clarifications or enhancements that assign straddle claims squarely to one of them [11].
A second mechanism runs through underwriting itself. Underwriters are now asking for bias testing records, human-in-the-loop protocols, model cards and evaluation results [12]. Those answers get incorporated into the policy by reference, which is the step that turns a description of your governance into something a carrier can use against you at claim time [13]. Halbreich's drafting answer is to require knowledge or intent and materiality before a misstatement bites, and to limit the consequences to the covered individuals responsible for it [14]. Read that against who actually writes the answers: the bias testing record comes from the ML team, and the human-in-the-loop protocol comes from whoever owns the workflow. The interview treats sign-off authority on AI use questions as an open question [15].
On the slow failure, the material is thinner than the topic deserves. Help Net Security's summary says Halbreich addresses when the claim clock starts for slow-building model degradation, and how business interruption coverage applies to outages at cloud and compute vendors [15]. The answers carried in the published excerpt develop straddle claims and the warranty language. The degradation trigger remains a question a practitioner has raised, without a settled position behind it yet.
Scope decides how much of this you should care about. If your company is not buying or being bought, straddle claims are somebody else's problem. The application-as-warranty problem arrives at every renewal [12][13], and by the interviewer's account the first call to coverage counsel usually comes after a demand letter has landed [16]. Dates and paperwork fixed long before any incident decide the bill, and they sit in documents most security teams never open.
