Curative says its AI agent has signed 23,000 health-provider contracts
A new Anthropic case study details automated negotiations, human approval for unusual rates, and a contracting team that grew rather than shrank.
Loading page…
A new Anthropic case study details automated negotiations, human approval for unusual rates, and a contracting team that grew rather than shrank.
Listen to this story
Curative’s contracting operation now uses agents for provider outreach, rate checks, contract edits, and follow-up, with human review retained for pricing exceptions. The company reports faster contracting and higher employee productivity, but the figures come from an Anthropic-published customer case study and have not been independently verified. Its approach pairs task-specific agents with escalation rules and message screening; Curative plans to apply agents next to care navigation.
Curative says adding fax support more than doubled Gwen’s win rate, including for practices that return handwritten contract edits.
A Claude Sonnet reviewer clears about two-thirds of requests above standard rates; uncleared exceptions go to CEO Fred Turner.
Curative says its contracting team has grown from 45 to about 55 people, with each employee roughly 2.5 times as productive.
Health insurer Curative says its AI contracting agent now signs 2,000–2,400 provider agreements a week—roughly matching its entire 2025 total. In a November 10 customer case study published by Anthropic, the company describes how Gwen negotiates with medical practices, while other agents check rates and messages and humans approve unusual pricing.
Gwen has signed 23,000 provider contracts, Curative says. These are company-reported results published by its AI supplier, not independently verified performance measurements.
Each Gwen run handles one provider, mostly using Claude Opus, and keeps no memory across deals. For incoming website inquiries, it researches the practice and proposes rates using purchased claims data and insurers’ published price-transparency files. It can show those market rates to providers during negotiations.
Gwen checks requested contract edits against Curative’s policy document, then writes Python code to merge changes. Curative says form submission to signature can take as little as an hour. Separate agents then check provider licenses and load contract data for claims pricing.
Curative says adding fax more than doubled Gwen’s win rate. Some providers handwrite changes on printed documents and fax them back.
Gwen does not have unrestricted pricing authority. Requests to exceed standard rates need an explanation for a Claude Sonnet reviewer. It clears about two-thirds; non-standard cases it does not clear go to CEO Fred Turner. Curative loosened its original human approval requirements in stages.
A Claude Haiku checker reads every outbound message and blocks those needing revision. Curative added it after Gwen copied providers on internal notes revealing how high it could go on rates. Blocked messages return with a reason so Gwen can rewrite them.
Curative initially barred Gwen from protected health information, keeping early mistakes in lower-risk work. Where agents now touch that information, back-end controls restrict what enters conversations, and activity is logged for audits.
The bar is, if you had a thousand people signing contracts, how many mistakes would they make that you regret?
Isaac Turner, Curative CTO, on evaluating agent mistakes
Since Gwen launched, Curative’s contracting team has grown from 45 to about 55 people, each about 2.5 times as productive, the company says. Employees give Gwen cold outreach and paperwork, then handle warm leads by phone. They receive compensation on Gwen’s deals too.
Curative also reports that human-agent teams send customized claims-dispute responses in under 30 minutes by email or fax, down from about 90 days. That measures response speed, not dispute resolution. Next, it plans to extend agents into care navigation, supporting the 130 people who guide and coach members.
Loading discussion...
Join the conversation
Explain whether automation should face a higher standard.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.