Key points
- OpenAI's tools handle 80-90% of Travelers claims intake and give Oscar Health adjusters 50% more capacity.
- Bastiaan de Goei says underwriters ignore up to 50% of broker submissions, a task OpenAI's Dots can do overnight.
- De Goei advises testing Dots personally and warns insurers that being regulated is no reason to wait.
OpenAI's Insurance Agents Are Already Working Claims
OpenAI's work with insurers has moved from pilots into production, and claims is where it shows. Travelers lets an OpenAI tool handle 80 to 90% of claims intake and Oscar Health halved verification time, but most insurers still automate single steps rather than whole processes. Persistent agents can carry a commercial submission from a broker's email through pricing and portfolio checks to a drafted quotation, leaving the underwriter to orchestrate and sign off. The message to a regulated industry is to test the tools now, because an 18-month transformation will be obsolete before it ships.
Takeaways
- OpenAI tools already run at carrier scale. At Travelers, 80 to 90% of claims intake data is fully handled by an OpenAI-powered tool, and a co-developed solution at Oscar Health cut claims verification time by 50%, giving each adjuster 50% more capacity.
- Optimising single steps leaves gains stranded. Insurers get good at one isolated step such as data intake or validation and the improvement stays there. Persistent agents work end to end, and in commercial underwriting they can work an inbox overnight where underwriters drop up to 50% of broker submissions.
- Underwriters orchestrate but keep accountability. The quotation and the responsibility for it stay with the underwriter. The tooling has to show the data, assumptions and steps behind a decision, which is what the sources-by-paragraph feature in ChatGPT for financial services is built for.
- Persistent agents codify knowledge held in heads. Every Dots agent keeps a scratch pad that accumulates a firm's workflows and exceptions, turning personal knowledge into corporate knowledge. Bastiaan spent a day without opening Gmail, Slack, Sheets or presentation software, only orchestrating his agent.
- A failed AI test is a process failure. ROI has to stay front and centre, and when an organisation sees no value the fault usually lies in the process rather than the product. If every person is 15% more efficient, the organisational gain is large.
- Cyber defence splits into two access tiers. Daybreak Blue gives verified defenders 80 to 90% of what an insurer needs for cyber defence, while Daybreak Red, at a higher threshold, covers penetration testing and red teaming. AI cuts both ways, so the net benefit has to justify the risk.
The 90-day playbook
- Test a Dots agent personally. They are available in any pro licence of ChatGPT, so a leader can use one without a procurement cycle and see what it does to their own week.
- Apply what you learn to claims and underwriting, the two processes where the work is document-heavy and the results are measurable.