The problem
Every time someone joins a client company, leaves it or adds a dependant, the insurer has to be told. Those changes go out in batches, and each batch was prepared and sent by someone in operations, then followed up until the insurer confirmed it.
The work grew with every account we took on. More clients meant more batches, and more batches meant more people.
What the emails showed
Before building anything, I went through twelve months of endorsement email with ten insurers, about 2,000 messages, to see where the time actually went.
I expected file formats to be the blocker, since each insurer wants its own. They weren't. The time went on reminder loops: sending a batch, waiting, chasing, and chasing again until someone confirmed it. That changed what to automate first, and which insurers to start with.
Three ways to send
Insurers don't all work the same way, so there's a channel for each kind.
- Insurer APIInsurers with an API get batches directly from our system, with no email in between.Where the insurer has one
- AI email agentSends the batch, reads the reply and follows up when nobody answers. The model sorts and extracts; the rules decide what happens next.For insurers that work over email
- RPA botLogs into the insurer's portal, uploads the batch and checks back for the result, once a day.For a portal-only insurer
Measuring it
"How much is automated" sounds simple, but the first numbers moved in ways that didn't match what ops saw. I set out what the share counts, so it means the same thing every month.
The first upload of a company's members is a different job, and none of it was ever automated. Counting it would hide progress on the day-to-day changes.
Turning on a new insurer shouldn't rewrite past months. Each batch is judged against what was switched on the day it went out.
The definition was tested against about 6,700 real batches before the dashboard used it, so the number on screen matches what happened.
Who did what
Me
The email research, scoping what the agent needed before it went live, how automation is measured, and the dashboards ops use to track it.
The team
Engineers built the email agent and the RPA bot, and ran the insurer integrations.
Result
Every batch that sends itself is one less in an ops queue. The team spends its time on the exceptions that need judgement, and taking on more accounts doesn't mean hiring more people. Automated batches also go out as soon as they're ready, so changes reach the insurer sooner.
