Most claims automation demos are videos. A narrator, a happy path, a cut before anything gets awkward. Ours is an interactive walkthrough of our AI claims agent working one claim file on the real platform screens, and it is a guided path too, we chose what you see. We chose a claim that doesn’t clear. The agent catches a problem, names what’s missing, and hands the file to a person. That’s the part worth watching.

Walk through the demo here. It takes about two minutes, and you click through it yourself. This post is the annotated version, written the way a handler would talk you through the file.

The case

Jan de Vries reports his e-bike stolen in Amsterdam. The file is worth EUR 4,200 and it arrives the way theft claims actually arrive: four documents of mixed quality. A purchase invoice, the policy schedule, the claimant’s own message, and a police report. All figures on the demo screens belong to that demonstration file, not to a client deployment.

We picked this claim because it hides the thing automation demos usually avoid: a detail in one document that has to agree with a detail in another one for the story to hold.

Step 1: Four documents arrive

The tour opens on the case screen with the file as it came in: purchase_invoice.pdf, policy_schedule.pdf, claimant_message.txt, police_report.pdf.

If you’ve worked a claims queue you know what those four files normally cost you. The reading alone eats the first part of the morning, and you can’t say what happened, let alone what’s missing, until you’ve done it. Here the documents land as they would in production, uploaded by the claimant or collected from the systems you already run, and the agent starts on them immediately.

Step 2: The agent reads the whole file

From the police report alone the agent pulls nine fields, including the one that ends up deciding this claim: which lock was actually on the bike. The tour tells you to keep that detail in mind. It means it.

Each conclusion the agent draws is shown on its own card with a visible confidence score; the theft assessment in this file sits at 94%. A conclusion the agent isn’t confident about is not silently accepted. It turns into exactly the kind of doubt that gets routed to a person two steps later.

Step 3: The lock doesn’t match

The policy schedule declares an ABUS Granit X-Plus U-lock. The police report records a chain lock.

A handler catches that on a good day. On a busy one it slips through, because nobody rereads a policy schedule against page three of a police report while the queue is growing. The agent caught the mismatch on its own, raised it as a risk signal, and cited the exact pages, so you can open both documents and check the conclusion at its source in seconds. A reviewer, an auditor, or a regulator can do the same a year later.

Consistency is one of a family of checks the agent runs on every file: coverage against the policy clauses, legitimacy of the parties and the invoice against registries, authenticity screening of photos and invoices for signs of tampering (a fraud detection layer we offer in the same workflow), and the settlement calculation itself. The demo shows you the check that bites in this particular file.

That citation trail is what separates an AI claims agent from optical character recognition (OCR) with a chat interface. Reading documents is table stakes. Noticing that two of them disagree, and showing you where, is the job.

One practical note: before the next step, the tour asks for a work email. That’s the gate between the walkthrough and watching how the claim resolves.

Step 4: The handler decides

The file doesn’t clear, and the demo doesn’t pretend otherwise. The policy wording is missing from the file, so the agent recommends requesting it before anything is paid out, and puts the claim in front of a person.

The handler’s screen shows four ways out: approve, deny, request info, or escalate, each with the agent’s rationale and the attached evidence. The agent does not own the decision. Your rules and your people do. A clean file where every check passes can settle straight through; this one has an open question, so it doesn’t, and that routing is the guardrail working, not the demo failing.

For European carriers this screen is also the compliance story. The European Insurance and Occupational Pensions Authority (EIOPA) published its Opinion on AI governance and risk management in August 2025, expecting insurers using AI in claims to demonstrate human oversight, explainability, and controls across the AI lifecycle under Solvency II and IDD. The handler-decision screen in this step is what that expectation looks like as a working interface rather than a governance slide: a person, four options, and the evidence trail attached to each one.

Step 5: Drilled on your own book first

The last stop leaves the claim and shows the Agent Gym. On screen, an orchestrator called Bike Claims Use-case is being built from the same four documents you just watched the agent work. Before an agent touches production it replays past claims and drills edge cases from your own book, and it goes live after your sign-off, not before. Ask any vendor how their agent is tested before it meets a live claim, and who signs off.

Why walk through this now

Two reasons, one operational and one about timing.

Operational: the gap between manual review and autonomous claims processing is not subtle. Files that spend days in email ping-pong between handler, claimant, and expert can clear in under an hour when intake, extraction, and checks run straight through.

Timing: most carriers have not built this yet, and the ones that move first keep the advantage for a while. WTW’s 2026 Advanced Analytics Survey of 59 property and casualty insurers in the United States and Canada found that only 33% use advanced analytics for fraud detection and 29% for severity assessment today, while respondents expect those figures to reach 65% to 70% within two years. The window in which claims automation is a differentiator rather than table stakes is measured in quarters.

The checklist: six things to look for in any claims-AI demo

1. Missing-document behaviour: when something is absent (here, the policy wording), does the system name it and recommend the request, or does a handler discover the gap later?

2. Visible confidence: does every conclusion carry a score you can see, and what happens when the agent is not confident?

3. Citation trail: does every conclusion point to its source document and page?

4. Cross-document consistency: does the demo include a detail that must agree across two documents, and does the system check it?

5. The guardrail: who defines the approve, review, and deny routes, you or the vendor’s model?

6. Pre-deployment testing: how is the agent drilled before go-live, and is go-live gated on your sign-off?

If a walkthrough can’t show you all six, the missing ones are where your handlers will be doing invisible cleanup work after deployment.

FAQ

What is an AI claims agent?

Software that processes an insurance claim the way a handler would: it reads the full file, extracts and scores what matters, runs coverage, consistency, legitimacy, and settlement checks with cited sources, and routes the outcome according to rules the carrier sets.

Does an AI claims agent replace claims handlers?

No. It removes the reading, checking, and chasing that should never have required human judgment, so handlers spend their time on the claims that do. The deny and review routes in the guardrail always end at a person.

How does the agent decide which claims settle straight through?

It does not decide the routing policy; the carrier does. A clean file where every check passes can be approved automatically, and anything uncertain or failed goes to a person. The thresholds and rules are configuration, not model behaviour.

What happens when documents are missing or unreadable?

The agent names what is missing and recommends requesting it before anything is paid out, as it does with the policy wording in the demo. Conclusions the agent is not confident in are flagged rather than silently accepted, and a person makes the call.

Does this meet EU expectations for AI in insurance?

EIOPA’s August 2025 Opinion on AI governance expects insurers using AI in claims to show human oversight, explainability, and lifecycle controls under Solvency II and IDD. A carrier-configured guardrail and a per-decision citation trail are direct answers to those expectations; your compliance team should still map them to your own governance framework.

Can I see the product without talking to sales?

Yes. The interactive walkthrough is public and self-guided. When you want to see it on your own claim types, request a demo.