E very major shift in claims over the last thirty years has been a change of medium. Paper gave way to fax, then email, then ticketing systems. What none of those transitions touched was the underlying shape of the work: a loss event shows up, gets in line, and a human moves it through stages while coverage policy lives in PDFs and tribal memory. Throughput is the scoreboard, and the real judgment sits entirely outside the software, in the head of the adjuster.
The queue has been the true operating system of claims all along. Everything else, from OCR and rule engines to chatbots and dashboards, was just built around it. And after years of adding accessories, the cracks are starting to show.
Speed and accuracy were never actually the tension
The industry usually frames the hard choice as speed versus accuracy. Push too fast and you pay on claims you shouldn’t. Move too slowly and your dealer calls someone else. Customers have always wanted both, obviously. What’s changed is that the tradeoff between them isn’t inherent anymore. It’s architectural.
Klarna is the cleanest cautionary tale in SaaS on this point.
In early 2024, they announced their AI assistant was doing the work of 700 agents. A year later they were publicly recalibrating and reinvesting in human support. “AI gives us speed. Talent gives us empathy,” a spokesperson said, walking the original framing back. The lesson isn’t that AI failed at Klarna. It’s that speed poured on top of a reactive system just produces faster bad decisions. The ceiling was the architecture underneath, not the model.
The companies getting this right aren’t bolting AI onto a queue.
They’re redesigning around agents. Intercom’s Fin reportedly runs fifteen different sub-processes to answer a single support question; parsing intent, pulling account state, choosing a reasoning path, confirming resolution. Their public framing for where this is all heading is the part worth borrowing: an agentic system needs roles it can fluidly move between, goals it can pursue, and memory that persists across the customer lifecycle. That’s the shape claims should be building toward too.
A few weeks ago, I wrote here that the fastest-moving F&I administrators have stopped treating AI as a feature and started treating it as an operating model. That argument still stands. What I want to add now is the layer underneath it because the operating model itself is only as good as the foundation it’s running on.
What the foundation must do
An agentic claims system isn’t a chatbot with a claims skin. It’s software given goals — adjudicate this VSC claim within policy, recover subrogation on this third-party liability, validate this repair estimate — along with the tools, the memory, and the bounded authority to pursue them. Humans don’t disappear. They move up the stack, from processing paperwork to supervising outcomes.
The infrastructure required to make that real is not small.
- Coverage logic — has to leave PDFs and enter a structured, versioned layer a model can reason over.
- Data — FNOL, contract, inspection, and payment data must live in one fabric, not five brittle integrations pretending to be one.
- Agents must execute — issue ACH, request documentation, dispatch an inspection, with full audit trails, not just describe what should happen.
- Every autonomous action — needs a confidence threshold, a reversibility of posture, and a clean human escalation path. No exception.
This is foundational work, and a thin wrapper over a frontier model can’t fake it.
The claims layer stops being an archive
Once the system can reason, the organization reshapes around it. Experienced claims professionals don’t shrink in importance; their leverage multiplies. One of them supervising a fleet of agents can cover volume that used to take a whole team, and they get to spend their time on cases where human judgment matters.
Product, actuarial, and operations start working from the same live signal, because the claims layer isn’t a system of record anymore. It tells you what’s happening in the portfolio in real time.
That’s the shift worth preparing for speed and accuracy together, because the architecture finally makes them compatible instead of competing.
A closing thought
The next two years in claims won’t be won by whoever has the most impressive AI demo. They’ll be won by whoever was willing to rebuild the foundation so that when the models get sharper — and they will get sharper, quickly — there’s a system capable of using them.
The queue was the ceiling. What replaces it is the floor.