Curative says its Claude agent signs 2,000 provider contracts a week

Curative, a US health insurer, says an agent it calls Gwen now signs 2,000 to 2,400 provider contracts a week, up from 2,100 in the whole of 2025. The figures are in a customer article Anthropic published on 10 November 2026, built on interviews with Curative's chief executive Fred Turner and chief technology officer Isaac Turner. Curative says Gwen has signed 23,000 contracts in total.

The article is promotional, but it is unusually specific about the controls around the agent and about what went wrong.

How the agents are arranged

• Gwen, mostly on Claude Opus. Each run handles one provider from first outreach to signature and keeps no memory across deals. It proposes rates from a purchased claims data set and the price-transparency files payers must publish.
• A rate reviewer on Claude Sonnet. When Gwen wants to exceed its standard rate template, it writes out why. The reviewer clears about two-thirds of these requests, Curative says, and anything non-standard goes to Fred Turner.
• A judge on Claude Haiku. It reads every communication before it is sent. A blocked message comes back with the reason, so Gwen can rewrite it.
• After signature, a separate credentialing agent checks the provider's licences, and other agents load the contract so claims can price against it.

Curative says each contract takes about 20 million tokens, roughly a tenth of the original cost. The article gives no dollar figure. From web form to signature can take as little as an hour, it says.

What broke

• The leak. Curative added the judge after Gwen copied providers on internal notes about how high it could go on rates.
• Specialised tools. "We started off with, 'Oh, we'll have tools,' and that doesn't work that well, because redlining is quite complex," Fred Turner says in the article. Gwen now has one tool, run Python, and does contract redlining with a docx library.
• Memory. Brad, a second Opus-based agent that sells plans through brokers, does keep memories. "You can end up making notes that are wrong and then get into these doom loops, where it has a bad memory that can never be cleared," Turner says.

What Curative says it learned

• Start with less authority than the agent can handle. Gwen began with a human approval queue for any rate above the template. "Having a second agent review rates worked better than just giving Gwen more authority," Turner says.
• Meet people on their channel. Curative says a fax tool more than doubled Gwen's win rate. Some providers reply only after the fourteenth email and the ninth fax.
• Improve the system from outside the run. Every few days Turner asks Claude Code to sample the last 10,000 runs, flag anything odd and open a pull request to fix it.
• Experiment away from protected health information first. Gwen was barred from it for a long time. Where agents touch it now, the article says back-end controls limit what enters a conversation and agent activity is logged for audit.

The contracting team grew from 45 to about 55 people after Gwen launched, and Curative says each is about 2.5 times as productive. Staff hand Gwen cold outreach and paperwork and take warm leads by phone. Curative pays the team on Gwen's deals too.

What the article leaves out

• Cost in dollars. Twenty million tokens per contract across three model tiers is not a price, and the article does not give the split.
• Error rates. There is no count of contracts signed at a wrong rate, messages the judge blocked or deals a person had to repair. Isaac Turner's stated bar is comparative: "if you had a thousand people signing contracts, how many mistakes would they make that you regret?"
• A control. The article compares Gwen's weekly rate with the 2025 total. It does not say how much of the gain comes from the agent rather than from a larger team or the fax channel.
• The provider's view. The article says faster contracting helps small practices get paid sooner. No provider is quoted.