The Broker-Carrier Gap
I spent two years inside the operations department of one of Europe's largest insurance carriers.
I measured the times. I sat next to the people who process policy changes and claims payouts. I watched how documents flow through the system. I counted the rework.
125 people. 1.8 million processes a year. Average lead time for a policy change: 64 to 80 days. Nearly one in five processes gets re-entered because the input was wrong or incomplete the first time.
The system they use crashes. The RPAs they built fail on decimals and incomplete documents. The average age of the operations team is 50. Average tenure: 23 years. There are no manuals. Knowledge lives in people's heads. And those people are starting to retire.
This is the largest insurer in the country. And the root cause of all of it is one thing: the entry point.
The entry point problem
Insurance carriers don't control how they receive work. Brokers, agents, and clients send documents in whatever format is convenient for them — emails with five attachments, PDFs, Excel files, photos of paper forms, free-text messages, scanned documents with handwriting.
The carrier's job is to parse all of that, re-enter it into their systems, validate it against business rules, and route it to the right team. Manually. This is the bottleneck. Not the underwriting. Not the final decision. The intake.
My team's conclusion, after working across six transformation projects at the same carrier: intake documentation is the universal bottleneck. It affects every product, every line of business, every macro process. The shape of the problem is always identical: unstructured input from one party needs to become structured data for another party. The variable format breaks every rule-based automation attempt.
Why RPA failed
Carriers have tried to automate this. They built RPAs — rule-based bots that follow deterministic paths. They fail because the input is too variable. An RPA expects a field in a specific location on a specific form. When a broker sends a photo of a different form with the field in a different place, the RPA breaks.
The automation ceiling for insurance operations is not compute. It is input variability.
Why agentic AI breaks through the ceiling
Large language models read documents the way a human would. They understand context. They extract meaning from a scanned PDF the same way they extract it from a structured JSON file.
The architecture: a broker sends an email with attachments → a classification agent identifies the request type → extraction agents process each document in parallel with per-field confidence scores → a validation engine checks against business rules → a confidence router decides: auto-approve (>90%), human review (70–90%), or manual queue (<70%) → a system writer pushes to the carrier's policy admin system via MCP → a feedback loop captures every human correction, improving future accuracy.
The key insight is the confidence router. You don't need 100% accuracy to create value. You need confidence-based routing. Human review goes from 12 minutes of data entry to 45 seconds of confirmation.
Day-30 results at Fidelidade: 84% of transactions auto-approve. Processing time: 43 seconds. Cost per transaction: €1.20, down from €18.50. Rework rate: 3.2%, down from 20%.
The BPO displacement nobody is talking about
Fidelidade pays a company called NewSpring €5,000 per month for 2 full-time employees to help with policy management. Those two people do exactly what relay does: open emails, read attachments, re-enter data, route cases. At approximately $500/month in API costs, relay replaces that work.
The global insurance BPO market is $68.4 billion (2026), growing at 6.36% CAGR. The real addressable market isn't just carrier operations automation ($12B). It includes BPO displacement ($68.4B). The real TAM is $80B+.
Why now. Why Europe.
Three things are converging. AI tooling crossed a threshold. The EU Platform Work Directive takes effect December 2, 2026, creating mandatory insurance demand for every gig worker in Europe. And the workforce crisis in carrier operations is structural: average team age 50, average tenure 23 years, no documentation.
Every well-funded competitor — Corgi ($108M), Harper ($47M) — is in the US. None of them are coming to Europe in the next 18–24 months. The regulatory cost is too high: IDD passporting, GDPR, ASF, DGSFP. The EU regulatory moat is an 18–24 month head start.
By the time a US competitor enters, IndieBacked has the carriers, the data, and the network effects.
What this is really about
This is not about insurance technology. Insurance is the domain. Every industry has a version of the entry point problem — unstructured input from one party needs to become structured data for another party. Healthcare has it. Legal has it. Government has it.
Insurance operations will be one of the first industries fully automated by agentic AI. Not because insurance is easy. Because the problem is perfectly shaped for what LLMs do best.
The team that solves it for insurance will have the architecture, the data, and the playbook to solve it everywhere. That is the real prize.