We spent six months building an AI workflow product with a commercial insurance design partner. We learned a lot. Then we pivoted. The underlying IP and workflows are still in service - now available to Claude as MCP tools rather than as a standalone product. This is a short note on why we changed direction.
The original bet
Commercial insurance runs on messy unstructured data. Submissions, schedules, policy wordings, broker emails, claim files, market correspondence - none of it structured, all of it consequential. That looked ripe for AI. We started building a workflow product to help underwriters and brokers cut through it: ingestion, classification, summarisation, drafting, evidence-aware review.
Our design partner gave us deep ground truth. What works in production, what does not. What the governance team requires before something goes live. How a regulator thinks about evidence trails. How a broker actually decides which market to send a submission to.
But while we built, the ground kept shifting.
The application layer moved
Every few weeks, the frontier model providers shipped capability that ate a meaningful slice of what we had built. Longer context. Better document parsing. Native tool use. The Model Context Protocol. Then ChatGPT Apps. Then MCP apps on Claude.ai.
The "workflow product" we were building was, more and more, a thin shell around a model that could increasingly do the workflow job itself. Platforms eat applications. The honest question is what is left when they do.
What doesn't get eaten
Four things, we think, do not disappear when the platform absorbs the application layer:
- Industry context. Knowing how a renewal pack is actually assembled. What an underwriter looks for in a submission. How complaints work under Fair Conduct. How a broker decides which market to approach. The model can read your documents. It cannot run your business.
- Implementation rails. Connectors to broking platforms, policy admin systems, claim systems, document stores. MCP servers built for the specific data shape of a regulated insurer. Identity, permissions, approval workflows, audit logs.
- Governance discipline. Source citations, role-based access, evidence trails, conduct supervision, exception queues, human approval before client-send. Operating disciplines, not model features. They live in the implementation layer.
- Change management. The unglamorous human work of moving a regulated team from "we have an AI pilot" to "we run this in production every day" - sponsorship, training, exception handling, feedback loops, role redesign. The model does not do this. People do.
So we stopped trying to be the application, and started being the team that puts the application together for the specific shape of a regulated business.
What changed in practice
We kept the industry context, the design-partner relationship and the technical team. The IP and workflows we built remain in service, now available to Claude as MCP tools rather than as a standalone product. We rebuilt the commercial offer around a fixed-price two-week Readiness Sprint, a workflow implementation practice, a governance practice, and accelerators for the workflows we know well.
We also stopped pretending we know every vertical equally. Insurance and brokerage are where we have gone deepest. Wealth and advice is the vertical we have moved into next. Banking and accountants we will get to when the right partners are ready to work alongside us.
Honest reflection
Six months is a short distance to walk before changing direction. But the moat was eroding visibly, and our design partner's real problem (how do we safely get AI into our operations?) turned out to be a bigger and more durable problem than the one our product was solving.
The pivot is not finished. The shape of the implementation layer is still being worked out. But the question (how does a regulated business put a frontier AI into production, safely, with evidence?) is the right question to be working on.
That is what Airclerk does now.