Open any insurance conference agenda and count the vendor booths. Underwriting copilots. Claims triage assistants. Broker email summarizers. Policy document extractors. Each one solves a real problem. Each one also asks your organization to sign a contract, grant data access, and trust a small team you've never met with a slice of your business logic.
This is the current shape of enterprise AI adoption, and I think it is structurally wrong.
The pitch from these startups is reasonable on its face: we specialize, we move fast, we've already built what would take your team a year. But specialization at the point-solution level does not require a new vendor. It requires better instructions to a model your organization already has access to. What these startups are actually selling is a prompt, wrapped in a UI, offered as a subscription.
The costs of this pattern are not hypothetical. Ten AI vendors means ten security reviews, ten data processing agreements, ten places your claims data or member records can leak. It means ten renewal negotiations and ten pricing models to track, none of which get cheaper as usage scales the way internal tooling would. It means the logic that decides how a claim gets triaged or how a policy gets underwritten now lives partly in your systems and partly in a vendor's black box, updated on their schedule, not yours.
Worse, it means nobody inside the organization fully understands how the work actually gets done anymore. Ask a claims manager to explain the end-to-end decision logic across five AI-assisted steps in a workflow, and you'll get two different vendor names and a shrug. I don't want to use too much jargon, but in a sense the organization essentially gets dumber over time with this approach.
Here's what the vendor-patchwork model misses: the entire point of large language models is that they are general-purpose reasoning engines that can be shaped, through careful instruction, into whatever specific task you need. That shaping, the actual encoding of how your organization underwrites a policy or adjudicates a claim or reviews a contract, is where the value sits. It is your domain expertise, translated into a form a machine can execute. That translation is intellectual property. It should not be handed to a third party and rented back to you monthly.
This isn't an argument for building everything in-house. Nobody should be standing up their own CRM or payroll system to make a point about self-reliance. The distinction that matters isn't how much effort something takes, it's whether that effort encodes your organization's own judgment. Commodity infrastructure, the CRM, the email platform, the cloud provider, is worth buying regardless of how hard it was to build, because none of it is where your advantage lives. The instructions that tell a model how your underwriters think, how your claims team spots fraud, how your actuaries reserve for IBNR, are a different category entirely. That's not infrastructure or background tooling. That's the business.
Organizations that outsource this shaping work are giving away the one asset that actually compounds. A vendor's model gets better for all their customers, including your competitors. An internally owned set of instructions gets better only for you, refined against your own edge cases, your own regulatory environment, your own institutional judgment about what counts as a red flag. One quadrant is a moat. The other is rent.
The startups selling narrow AI wrappers around insurance workflows aren't going away, and some of them will genuinely help in the short term. But five years from now, the organizations in the strongest position won't be the ones with the most AI vendors. They'll be the ones who kept their own judgment in their own hands, and simply gave it a new way to execute.
Finally, this isn't a post against AI startups. There are enormously useful ideas out there that should become great companies that help enterprises. However, as is the case with big technological shifts, there are a ton of quick-to-market LLM-powered startups that are marketing exactly the type of solution that I describe in this post. Stay safe out there!