Generative AI can surface assumptions and challenge big plans, but insurance teams need to check its knowledge and resist answers that simply agree.
Generative AI is starting to offer strategy teams a fast way to question plans before they become costly commitments. For insurers, whose products and operations depend on specialized market knowledge, it may be most useful as a tireless challenger: a tool that surfaces assumptions, asks what could fail and pushes executives to examine the case for doing nothing as well as the case for acting.
Bill Gates has described using several AI systems in the Gates Foundation’s annual strategy discussions about how to allocate its next $10 billion. In some sessions, the systems are told to wait until someone asks them a question. In others, they can interrupt when they hear a claim or figure they think deserves scrutiny. Gates said the systems could become peers in complex discussions, contributing without making final decisions.
That’s an ambitious use. Paul Carroll, drawing on research into failed business strategies, sees a more measured role for now: give an AI a plan, then question it repeatedly and insist that it support its reasoning with evidence. The goal is to make room for doubts that can get lost when a senior leader already favors an idea.
Give the plan a skeptical review
Carroll’s starting point comes from a former boss, Mel Bergstein, who summarized strategy as “what you don’t do.” In Carroll and Chunka Mui’s book Billion Dollar Lessons, researchers examined 2,500 major failures over two years. Their proposed safeguard was a “devil’s advocate” who identifies the assumptions underneath a plan and tests the risks executives may be tempted to overlook.
An AI can take on part of that questioning role. A team considering an acquisition might begin with a broad prompt about what could go wrong, then follow up with focused questions: How might competitors respond? What happens if borrowing costs rise or the economy weakens? Can the buyer actually sell more products to existing customers? Are the expected efficiencies realistic? Which comparable efforts ran into trouble, and why?
The same process applies to plans beyond acquisitions. Teams could test a new line of business, a major software replacement or the adoption of customer service chatbots. They can also ask what risks come with leaving a proposed change on the shelf. That last question matters because strategic reviews often spend more time building a case for action than weighing the cost of standing still.
AI’s value here depends on the quality of the exchange. A team should ask for the evidence behind a claim, probe each assumption and tell the system to be direct. A polished response can still be weak if it merely restates the plan in confident language.
Why independent challenge matters
Large companies already commission extensive analysis, and acquisition teams may seek advice from investment banks. But Carroll points to incentives that can tilt those reviews toward making a proposal work. Employees may focus on supporting an idea they believe the chief executive favors. An adviser paid when a deal closes may have less reason to argue against it.
Carroll and Mui saw this dynamic while reviewing a company’s proposed move into autonomous vehicles. Consulting studies had portrayed the robotaxi business as a clear success for the first entrant. Yet employees working on the project knew about operating and safety concerns that later slowed the business’s rollout. Those doubts came out in interviews for the authors’ independent review.
Warren Buffett used a different way to seek competing arguments on potential acquisitions: he set aside equal sums for one bank to argue for a deal and another to argue against it. The bank whose case prevailed earned the fees; the other did not. The arrangement gave both sides a reason to build their arguments carefully. An AI system doesn’t have the same fee incentive, but it can still produce answers that lean toward pleasing the person asking the question. That tendency needs to be tested.
Insurance needs market specific input
Carroll is cautious about treating AI as a full participant in complex strategy sessions, especially in a business such as insurance. He says teams would need to test whether a model understands the details of the market and provide information specific to their company and area of work, including applicable rules. Publicly available material may not give a system the context needed to assess a specialized insurance decision.
That caution is particularly relevant when a plan involves products, claims operations, pricing or technology. A model can help organize questions, but the answers require review by people who understand the business and can check its underlying evidence. Teams should also separate a useful challenge from a confident guess: the system’s fluency doesn’t establish that its account of a market condition or operating risk is correct.
Small assumptions can drive large failures
The case for scrutiny is plain in Carroll’s examples of companies moving beyond what they knew. A leading cement maker treated its presence in home construction as evidence that it belonged in household products, then expanded into furnaces and lawn mowers. It later went bankrupt and was bought by a competitor. Avon identified its caring culture as its central strength and bought businesses in medical equipment and retirement homes, then sold them at a major loss after struggling to run them.
FedEx, meanwhile, spent hundreds of millions of dollars building a document delivery service that relied on drivers carrying papers to and from fax machines. The company failed to account for how quickly those machines would become affordable enough for businesses to own their own. In each case, the move rested on a premise that deserved harder questions before the company committed resources.
Carroll and Mui estimated that 47% of the failures they studied involved strategies that should not have survived scrutiny in the planning room. In his view, AI won’t catch every bad idea, and it will take time to learn how to use it well. Still, a system that helps teams expose weak assumptions could improve the odds of spotting trouble early. For insurance leaders, the practical first step is to use AI to sharpen the questions, supply the market context it lacks and verify its answers before relying on them.
Paul Carroll writes that generative AI can help strategy teams identify assumptions and consider what not to do.
