The situation
A commercial insurance agency was spending up to three hours to quote a single customer. The work itself was not complicated. It was the same intake information, typed by hand into a dozen different carrier portals, one after another, with the formats and field names slightly different every time.
Every hour spent re-keying is an hour not spent on the parts of the job that need an underwriter’s judgment. The agency knew this. The question was whether anything could be done about it.
What was in the way
Commercial insurance is a difficult place to automate, for real reasons.
Most carriers have no API. Several enforce two-factor login, which stops most automation approaches at the door. The industry appetite finder, the tool meant to say which carriers would even want a given risk, was wrong in both directions: it suggested carriers who would decline, and it missed carriers who would have quoted.
And the failure modes are permanent. A slip with carrier credentials or a premature bind cannot be taken back. No off-the-shelf tool could work the portals safely under those conditions.
What we did
We built it, with the constraints treated as design requirements rather than obstacles to route around.
- Schema-driven intake. One structured record of the risk, mapped to each carrier’s own field names, so the information is entered once.
- Rules preflight. Appetite and eligibility checked before an agent opens a portal, so time is not spent on carriers who will decline.
- One browser agent per carrier, running in parallel. The dozen portals are worked at once rather than in sequence.
- Two-factor pause and resume. When a carrier challenges, the agent halts and waits for a human to clear it, then continues from where it stopped.
- Human approval before bind. No agent binds coverage. The irreversible step stays with a person, every time.
- Full audit trail. Every action, every field value, every approval, recorded and reviewable.
Routine steps that need no reasoning run as deterministic code rather than model calls, which is why the running cost of the repetitive portion is effectively nothing.
What came out of it
- Three hours of manual re-keying designed down to minutes
- A working prototype, backed by 90 automated tests
- The routine steps running as code at zero token cost
What this means for a growing business
Two things are worth taking from this one.
The first is that “there is no API” does not end the conversation. A browser agent can work a system that was only ever built for a human, which puts a large amount of previously untouchable work in scope.
The second is that agents in a regulated business are a design problem before they are a technology problem. The value is in the parallel execution. The safety is in deciding, deliberately, which steps a human still has to sign. Bind approval and the audit trail are not overhead bolted on at the end. They are what makes the rest of it usable.
When the right recommendation is software, we build it, with the human approval points and audit trail a regulated business needs.