Industry
AI employees in insurance
Tobiloba Odejinmi · 29 Jan 2026 · 6 min · 944 words
Direct answer
An AI employee in insurance does the first pass: reads the onboarding pack, pulls the fields, answers the broker question your team already answers, and leaves claims judgment with a person. We grew Insurpass to more than 100 companies and a team of 17, then sold it. Buyers did not ask about slogans. They asked whether the product would still run after the wire hit.
- Paper to digital is still the job. The employee reads the pack. A person still binds the risk.
- Broker support and incomplete onboarding are better first processes than 'AI claims'.
- If you cannot show a log, you do not have a system. You have a trick.
- Diligence will ask cost, uptime, and who can explain the database. Build for that now.
Insurance is still a pile of documents
Insurpass existed because too many insurance products still launched on paper. We grew it to more than 100 companies, led a team of 17, and sold the company. The work in the middle was not a slogan. It was getting a product online so someone could buy it without a courier.
An AI employee does not change that job. It changes how many times a person has to read the same pack. Pull the fields. Flag the gaps. Leave the bind for someone whose name is on the file.
First pass, not the claim
I get asked about 'AI claims' more than anything else in this industry. I treat that question the way I treat a silent credit decision. The intake can be automatic. The payment cannot be a vibe.
The safer first jobs are the ones insurance already hates: incomplete broker onboarding, the same policy questions, the endorsement that is waiting on one missing page. Those have a shape you can check. A contested claim does not.
- Read the onboarding pack and list what is still missing.
- Answer the broker questions the team already has a script for.
- Draft a claim summary. Do not issue the payment.
- Write every step to a log a supervisor can open without you.
The old database was the real product risk
At Insurpass we also moved the old database work onto a faster stack. Buyers notice that kind of thing. They notice it more than they notice a demo that extracts a field from a PDF.
If your AI employee writes into a store only one person understands, you have made the company harder to sell and harder to run. Put the output in the same systems the ops team already trusts. Then keep a person who can explain the schema.
What a buyer opens first
When we sold, the conversations that mattered were not about vision. They were about whether the product would still run after the wire hit. Cost. Uptime. Who can change it without the founder in the room.
Add agents later and the same questions apply. Who owns the workflow. What it costs when volume doubles. Whether you can turn it off. Whether a miss is explainable. Write those answers down before someone asks.
Keep the bind with a person
Insurance is regulated for a reason. A model that is usually right is not a control. A named underwriter or claims lead is a control. The employee can make that person faster. It should not make them optional.
If that feels conservative, good. Conservative is how you still have a company when a supervisor wants the file.
Build one workflow you can hand over
Pick the pack that burns a week every month. Map it. Connect to the tools the team already uses. Go live with a write-up and monitoring. Then stop talking about transformation until that loop has a miss rate you can say out loud.
What I would tell a new insurance ops lead
Do not start with claims bravado. Start with the pack three people already read in the same order. Write the steps on one page. Mark the cases a person still has to see. That page is the spec.
Then ask who will own the miss at 2am. If the answer is 'the model', you do not have an owner. If the answer is a name, you can build. We led a team of 17 at Insurpass. The people who made the company saleable were the ones who could explain a file, not the ones who could describe a vision.
I still write the code. If you have one insurance process that still runs on people copying fields from a PDF, that is usually the first job. Thirty minutes is enough to see whether it is worth a week.
Questions people ask
Did Insurpass run on agents when you sold it?
No. We sold a product companies actually ran. The lesson still applies: if you later put AI employees on that stack, the buyer will ask who owns the workflow, what happens when it is wrong, and where the logs live.
Can the employee decide a claim?
It can intake the claim, list missing documents, and draft a summary. A named adjuster still owns the outcome. Silent claim decisions are how you buy a regulatory event.
What should an insurer automate first?
The onboarding pack that three people read in the same order every week, and the broker questions that never change. That pile has a shape. Claims bravado does not.
How do you keep it explainable?
Structured output. An audit trail. A person on the bind and the pay. If a supervisor cannot replay why a file moved, you are not ready.
Will a buyer care about this?
Yes. Diligence is a tour of the boring systems. Agents without owners look like magic. Magic looks like risk.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
