I’m building Hammer Labs (https://hammer.ai) to study when healthcare AI agents are wrong and when they should refuse to answer (judgement).

It started after spending 15 years building AI for insurers, hospitals, data companies, and startups. Almost every system ended with "a human reviews the output". That person was usually a nurse, medical director, or certified coder. These are some of the hardest people to hire, and the same people automation was supposed to help.

The problem is that real claims do not have an answer key. You cannot reduce human review until you can measure when an agent is wrong.

Getting claims data is also difficult. It can take a year of data agreements, privacy reviews, and procurement. Even then, you may not know what the correct decision should have been. So we generate claims. Utilization and case mix come from published data. Claims are priced using real fee schedules and contract terms. Payers behave differently, like real payers do. We plant errors on purpose, so the correct answer exists before any model runs.

On top of that, we are building benchmarks for overreach, refusal, errors by record type, and detection time. We are also building small MCP tools that refuse when evidence is missing. Every number includes its source, date, and basis.

What I find interesting is how much of this sits between actuarial work and machine learning. Both are needed, but I do not see many people connecting them.

25 published refusals: https://hammer.ai/worlds/refusals/ .

Runs on rate and policy evidence https://hammer.ai/reimbursement-evidence/ and savings claims https://hammer.ai/savings-claims/ .

AgentPlugin is Apache-2.0: https://github.com/hmmrlabs/hammer-plugin