What happens when carriers feed bad reserving data into AI?

Interest in agentic AI within insurance, specifically claims and underwriting are two functions analysts most consistently point to that are leading adoption.

A Celent survey found that 22% of participating carriers expect to have an agentic AI solution in production by the end of this year. While reserving hasn't drawn the same attention yet, it's starting to surface in industry conversations as the next area under consideration. The idea being that AI can dynamically optimize how much capital sits in reserve, reducing trapped cash and freeing funds for better deployment. 

However, Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls. It's important to ask what a near-40% cancellation rate could mean if agentic AI were applied to a function like reserving, where mistakes don't announce themselves. A bad reserving model can sit quietly for months before anyone catches that the number was wrong, and that raises a harder question than whether agentic AI belongs in reserving. 

For carriers that operate through delegated authority structures — relying on TPAs, MGAs, and coverholders to handle claims — the data picture that an AI reserving model needs to function is fragmented, delayed and often incomplete. This isn't an argument against agentic AI in reserving. It's an argument that the enthusiasm around it is running ahead of a harder conversation about the conditions under which it actually works, and what needs to be true before the promise becomes operational reality. 

Why reserving is already complicated

There are three recurring problems with reserving in delegated authority structures. The bordereau lag gets the most attention, but it's actually the easiest to solve. Two harder ones sit underneath it. 

The first is definitional. Carriers, TPAs and coverholders don't always define reserve the same way, or update it on the same schedule. If the bordereau lag is consistent and predictable, actuaries account for this using standard incurred but not reported development, estimating how much unreported activity a known delay is likely masking. But no adjustment fixes a dataset where reserve means something slightly different depending on which entity reported the number. 

The second is structural. Policy administration and claims management systems have run insurance operations for decades as isolated silos. Reserve-setting, indemnity payments, subrogation recovery, and reinsurance recoveries often live in separate systems that were never built to reconcile with each other, let alone in real time. A model trying to flag reserve adequacy is working from whatever fragments those systems happen to share, not from one live picture of the claim. 

Both problems predate AI. An AI reserving model doesn't create them, but it does depend on them being solved first, and in most delegated authority structures, they aren't. 

What agentic AI needs to work

The technology itself is frequently layered on top of existing platforms rather than built into the workflow, which creates operational friction and inconsistent data. For fund management specifically, that's a real risk. A model sitting outside the system of record can flag a problem, but if it can't act inside the same claims-and-payment workflow as the fund movement itself, a gap opens between the moment something is flagged and the moment anyone can act on it. That gap is exactly where fraud and reserving errors live. 

There's a precedent for this outside insurance. Earlier in my career, I ran into this directly. Banks' core systems, the central ledgers that process transactions and settlement, were built for end-of-day batch processing, not continuous data. Adding the real-time data that AI-driven fraud detection or credit decisions require meant working against the grain of that infrastructure, not with it. 

The failure modes

Bad reserving data run through an AI model doesn't fail loudly, and that's what makes it dangerous. Thin or stale data can understate true exposure, leading to under-reserving. Just as easily, a model can overcorrect: defaulting to caution because it can't verify what it's seeing, leaving capital sitting idle that didn't need to. The quieter failure is the hardest to catch: numbers that look internally consistent and confident, but are wrong, because no one told the model the data underneath had never been harmonized in the first place. 

The questions carriers should be asking

None of this is an argument against agentic AI in reserving. It's an argument for answering a harder set of questions first. Questions I haven't heard settled with confidence in the conversations I'm having across the industry. Does reserve mean the same thing across every entity in the delegated authority chain feeding the model? Is the data the model sees pinned to a snapshot, or is it a moving target? Can the model act inside the same claims-and-payment workflow as the fund movement it's evaluating, or is there a lag between flag and action? 

These are all questions that are answerable now, and they're cheaper to answer now than after adoption. The enthusiasm around agentic AI in reserving isn't wrong. It's just ahead of the infrastructure required to earn it.