Distribution Executives Warn Against AI Overreliance Before Deployment
Three distribution executives put the same warning three different ways in remarks Digital Insurance reported on July 20: test the AI before it ships, because a bad launch cannot be taken back. Chubb’s Nicholas Davis, senior vice president and national distribution leader, said that “launching before testing properly is a mistake any organization could make. You can’t go back,” and that oversight and governance can help ensure the output is correct. His caution was about pace as much as controls: “You have to be mindful of how quick you move everything to a new system.”
Nabeel Tanveer, WTW’s head of affinity and programs for the U.S., drew the line by function rather than by technology. The tedious parts of broking are the ones he expects to go: “policy check-in, the billing, the maintenance. You don’t need an insurance license to do that duty. So AI is going to dominate that space.” Negotiations between carriers and brokers he put on the other side of that line, saying he does not expect them ever to be handed to an AI agent. His own department is cautious about taking AI-produced material to clients at all, on the grounds that catching a wrong variable takes someone with experience in the space: a model, as he put it, “is making assumptions like a human, and some of those assumptions are incorrect.”
Westfield Insurance’s head of distribution, Tom Zacharopoulos, speaking in a recent industry webinar, described the posture his company is taking as “trust and verify,” and said the discipline is validating “all the time that the information that we are utilizing is actually the best information.” Westfield intends to move slowly through the learning phase: “You can’t just jump into the deep end of the pool.”
None of the three used the language of compliance, and all three described a control an examiner would recognize. Question 8 of Exhibit C in the NAIC AI Systems Evaluation Tool asks a company to discuss how a model was validated prior to being deployed and how its performance is monitored on an ongoing basis, and question 9 asks for the last date it was tested. That is the paperwork version of what the executives are describing, and the word “prior” is doing the same work in both. An executive who says the review has to precede the launch is describing a record somebody can ask for later.
The checkpoint they describe is narrower than a full governance program. A named owner reads AI-generated client material before it leaves the building; an AI-assisted negotiation position gets a documented second set of eyes. Tanveer’s distinction is the useful one for sizing that control: heaviest where judgment and relationship carry the outcome, lightest where the task is structured and an error is cheap. It matches the contrast he drew between insurers reaching for AI on simpler products such as pet insurance and brokerages, even at middle-market level, still needing an agent in the conversation.
Sequencing is what separates a control from a cleanup. Review processes assembled after a failed launch are written under pressure, in the shadow of a client complaint or a regulator letter, and they read that way. Assembled first, they are just workflow, and they carry dates that run from before the first client saw anything, which is the record the governance framework implementation guide is built to produce. None of the three executives said how long their own review actually takes. That number is what decides whether the ordering holds when someone else has already set the launch date.