Agentic Adjusters
What changes when claims handlers stop being users of software and start being supervisors of it.
The single most under-discussed shift in enterprise software right now is what happens to the role of the human user when the agent gets good. In claims, that shift is happening faster than most operators realize, and the design implications are significant.
From user to supervisor
For thirty years, the adjuster has been the user of claims software. The software showed information. The adjuster made decisions. The software wrote the decision back. The interface optimized for getting data into and out of the human as quickly as possible.
Agentic claims software inverts this. The agent gathers context, synthesizes the file, drafts the next action, and proposes a decision. The adjuster’s job is no longer to make the decision in the first instance — it is to adjudicate the proposal. To accept, refine, override, or escalate.
This is a different job. It demands a different interface. And almost no one is designing for it correctly yet.
What supervisor UX requires
When you redesign for supervision rather than execution, three things have to change:
1. Density rises, but so does scannability.
A supervisor needs to see more at once — the agent’s reasoning, its sources, its confidence, its alternatives — but they need to parse it in seconds, not minutes. This is where most agentic UIs fail. They treat the agent’s output as the final artifact and the supervisor as a passive reviewer. In reality the supervisor is doing high-bandwidth pattern matching. The interface needs to support that.
2. Confidence becomes a first-class citizen.
Every recommendation the agent surfaces has to come with a calibrated read on how sure the model is. Not a vibe. Not a star rating. A number that’s been stress-tested against ground truth and that the supervisor learns to trust over time. When confidence is low, the UI should visibly hand more authority back to the human. When confidence is high, it should visibly compress the human’s role.
The cost of a wrong decision is not symmetric. The interface should not pretend it is.
3. The audit trail is a feature, not a footnote.
In a regulated, decision-heavy domain, the answer to “why did we pay this claim this way” is not optional. Agentic systems make this both easier (every action is logged) and harder (the action chain is now branching and probabilistic). The systems that win will treat reasoning provenance as a core product surface, not a compliance afterthought.
The honest failure modes
Three places this goes wrong if you’re not careful.
Automation bias. Supervisors who see confident-looking output start rubber-stamping it. The interface has to actively resist this — by deliberately surfacing doubt, by making override frictionless, by rotating spot-checks.
De-skilling. If the agent does the synthesis every time, the human loses the muscle to do it themselves. Over five years, you end up with a workforce that can’t function when the agent is wrong. The fix is intentional — periodic “agent off” cases, training paths that keep the underlying skill alive.
Accountability laundering. When a decision is made by a model and approved by a human in 800 milliseconds, who is actually responsible for it? If the answer is no one, the regulator will eventually answer it for you. Agentic systems need a clear, defensible theory of where authority lives at every moment.
Where I’m placing the bet
My read is that within three years the best claims operations will look more like high-frequency trading desks than call centers. A small number of highly skilled humans, supervising a large number of agents, with interfaces designed for the cadence and density of that supervision. The operations that don’t make this transition won’t disappear — they’ll just become structurally less profitable until they do.
The interesting design work — the work I’m spending my days on — is at the interface layer. Not the model layer, not the data layer. The thin, high-stakes membrane where the human meets the agent and decides what to do next. Get that right, and everything else compounds. Get it wrong, and nothing else matters.