ClinicalGate: the HIPAA-safe approval gate for agents that click through EHRs
The problem
OpenAI's computer-use agents now read screens and click through electronic health records — CVS, Cigna and Ambience are testing them — but every click an agent makes inside a patient record is a HIPAA event. A misclick can disclose PHI to the wrong chart, and retrospective compliance tools only discover it in next quarter's audit, after the fine clock started.
The idea
Why now
Computer-use agents for healthcare launched this week; Liquid d1.2 decision models now hear audio and answer in milliseconds; Sarvam Prism answers in 44ms; Privacyboard automates privacy compliance drafting. The gate is newly cheap enough to sit in front of every click.
What it combines
OpenAI computer-use agents for healthcare + Liquid d1.2 decision models + Sarvam Prism's 44ms verdicts + Privacyboard compliance automation. Agents alone click blindly, decision models alone score without context, compliance tools alone audit after the fact — together they are a live gate that approves, blocks, and documents every clinical action as it happens.
MVP
Weekend MVP: a proxy in front of one EHR test environment that logs every agent action, scores it with a fast decision model, and blocks writes to patient charts until a human taps approve in Slack. Deliberately skip: audio handling, multi-EHR support, and the full audit-report generator — ship the log and block first.
How it makes money
The moat
Vanta and Drata would have to rebuild from retrospective evidence collection into live, per-action clinical gating inside EHR sessions — a different architecture requiring clinical-workflow integration and per-action decision models they do not have. EHR vendors cannot credibly gate themselves: the auditor cannot be the party being audited. The per-hospital approval corpus is also a data advantage: every approved and blocked action makes the gate sharper for that hospital's workflows.
Distribution
B2B2C via EHR vendors and health-system IT — sell as the compliance add-on inside Epic/Cerner-style agent rollouts and through the agent platforms deploying into hospitals. Direct sale to health-system CIOs and compliance officers.
Why it wins
Vanta, Drata and Secureframe collect evidence after the fact; none of them can stop an agent mid-click. EHR vendors' own audit logs record what happened, not whether it should have. ClinicalGate is the only layer whose deliverable is a prevented disclosure, not a prettier audit report.
Risks
Biggest risk: hospitals will not let a startup sit between agents and patient records without a BAA and a security review that takes months. The MVP de-risks it by starting in a sandbox EHR with synthetic patients, proving the gate's precision there before any real PHI is in scope.
Build it with
- OpenAI computer-use agents for healthcare: agents that click through EHRs and payer portalsThe agents that click through EHRs and payer portals — the traffic the gate sits in front of.
- Liquid AI d1.2 and d1.2-ultra: decision models that now hear audioDecision models that now hear audio — the per-action scorer, including for voice-driven clinical workflows.
- Sarvam Prism44ms verdicts on every action so the approval gate never slows the clinician's agent down.
- Privacyboard: privacy compliance on autopilotPrivacy compliance on autopilot — the policy layer the gate enforces.
Repo to start from
clinicalgate — a proxy that scores every healthcare agent action against HIPAA policy in milliseconds, blocks risky writes for human approval, and emits the audit trail automatically.
Evidence
- OpenAI computer-use agents for healthcare (Fierce Healthcare)
- Liquid AI d1.2 decision models
- Privacyboard: privacy compliance on autopilot
- Vanta pricing 2026: Essentials $14,000/year
- HHS OCR HIPAA settlements: $8.33M collected in 2025, Solara $3M
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.