We Built an AI Agent to Fill Out Insurance Carrier Portals
How a Codex agent works carrier portals end to end and stops at a saved quote. An outline of the setup, with a polished version coming.
Heads up: this one is an outline, not a copy-paste build. We're turning it into a product, so the carrier files, rules and skill files stay private for now. Here's the architecture so you can see how it works.
What you'll build
A Codex agent that takes a new client from a master sheet, logs into carrier portals like Liberty Mutual and Travelers, fills in every screen and stops at a saved home and auto quote. It never buys anything. A licensed producer checks the quote before it reaches the client.
Before you start: this takes time
Nothing here works on the first run. Our first carrier took about an hour and 10 to 15 questions from the agent. Getting to roughly 90% accuracy took many supervised runs per carrier, with a licensed producer checking every quote. Expect the same.
We're working on a more polished version that skips most of that tuning, and we'll release it when it's ready.
What you'll need
- Codex (ChatGPT) with computer use
- A Google Sheet to act as the master client sheet
- Agent logins for the carrier portals you quote
- A licensed producer who checks every quote
Step 1: Split the work into two agents
An intake agent reads client emails and conversations and writes the details into the master sheet. A portal agent reads the master sheet and works the carrier portals.
Keep them apart. The portal agent never sees the inbox, and the intake agent never touches a portal. That's your privacy line.
Step 2: Build the master sheet
One row per client, with the fields the portals ask for: names, dates of birth, address, vehicles, drivers, current coverage. The portal agent only reads from here. Finished quotes go into a separate Run Log tab.
Step 3: Map each carrier, one at a time
Run the agent through one portal with the producer on call. Have it log every question it sees and every assumption it makes into a file for that carrier. Mark which answers are the same for every client and which need that client's data.
One portal had about 218 questions. Start on the strongest model you have, because it will ask a lot at first.
Step 4: Add a wrap step after every run
At the end of each client, the agent reviews the conversation: what it got wrong, what it had to ask, what the producer corrected. It writes those lessons back into its skill file. This is why run five goes smoother than run one.
Step 5: Check it against the producer
Have the producer quote the same client by hand. Any difference in the final number becomes a new rule for the agent.
Step 6: Track four numbers every run
- Autonomy: how many times it stops to ask a question
- Accuracy: does its quote match the producer's
- Usage: what each quote costs. Ours went from 20 to 30% of a $100 Codex plan per client on the top model to about 1% on a cheaper one, once the skill file was solid.
- Time: ours takes about 25 minutes against the producer's 12. That's fine, because nobody has to watch it.
Step 7: Keep a human on the last step
The agent stops at a saved quote. The producer checks coverages, reports and data, then presents it, takes payment and issues the policy. Never give the agent payment details.
On logins: most portals only ask for MFA once a week or once a month, so one signed-in session covers most runs.
Resources
- Follow the channel for the polished release: https://www.youtube.com/@austinpeechatt