Start a pilot
Put a named physician
behind your AI output.
Your model drafts the prior-auth determination, the letter, the coded note, the answer to the patient. The design: a licensed, registry-checked physician signs it, revises it, or declines it with a reason, and you get a receipt anyone can verify. The law increasingly requires the human. Next: a specialty-matched reviewer, named on the record; today: no reviewers are enrolled and the sandbox runs synthetic samples in your browser.
Pre-go-live pilot. The sandbox runs synthetic samples in your browser and stores nothing. A BAA is offered to qualifying covered entities.
Three steps. The first one takes two minutes.
Nothing here needs a call, an account, or a contract until you want patient data in the loop.
Run a sample through the physician layer
Pick the kind of output your AI produces, load a clean or flawed sample or paste your own synthetic one, and watch the review loop return a sign, revise, or decline with an attestation record. Runs in your browser; nothing is sent or stored.
Open the sandbox →Tell us what you ship
The output type, the decision it feeds, rough monthly volume, the states you operate in, and whether patient data is involved. That is enough for the founder to say yes, not yet, or here is what has to be true first.
The form is below →Review, then receipt
Next: a specialty-matched physician reviews each output, and each determination is hashed and verifiable at harnesshealth.ai/verify/<id>, so your compliance team can check it without trusting us. Today: the sandbox runs synthetic samples. Patient data waits for a BAA.
See how a receipt verifies →What every output is designed to get
- A named reviewer. A physician verified against the CMS NPPES registry, matched to the specialty of the output, bound to a signing key only they hold. Today: the sandbox runs synthetic samples.
- A real decision. Sign, revise, or decline — with a documented rationale.
- A receipt. NPI display, timestamp, document hash, and signature, verifiable by anyone at harnesshealth.ai/verify/<id>. Today: the only signer is a demonstration key.
- A closed gate. Designed so that outside a covered specialty the gate fails closed and tells you so. Not yet enforced in production; see /trust/hard-intercept.
Terms
Production terms are set in the pilot and disclosed before you accept. There is no checkout. How reviews are designed to work: clinicalswipe.com.
- Sandbox: nothing to buy, no account. Synthetic or de-identified outputs only.
- Production: terms set in the pilot, after a BAA.
- Never: a share of your revenue, a fee tied to approvals, or a fee tied to referrals.
Why a vendor needs this now
CMS's WISeR model, running from January 2026 in six states, requires that every non-affirmation be reviewed by a human clinician with relevant expertise and that no denial be made solely by technology. Seven states passed laws in 2026 saying the same thing to insurers: AI may draft or screen, but a licensed human makes and signs the adverse decision. If your product feeds one of those decisions, the human is now a requirement, not a feature.
Most of those laws bind the plan, not the vendor. A vendor selling into a plan supplies the clinician the plan must have — and the record that proves it. Whether a given arrangement satisfies a given statute is a question for the plan's counsel; we supply the named, licensed, specialty-matched reviewer and the receipt.
The other reason is cost. Mayo Clinic's chief AI implementation officer told Becker's (August 31, 2026) that the cost of maintaining AI systems is “far higher than I think any of us in the industry really thought,” and that engineers who can never let go of a live product become a big challenge. A signature per output turns the human part of that burden into a variable cost. The numbers →
The statutes, the dates, and the sources →·Built for a WISeR participant? →Straight answers
What do I get for each output?
By design, a specialty-matched, registry-checked physician signs, revises, or declines the output with a documented rationale, and the determination becomes a hash-anchored record anyone can verify at harnesshealth.ai/verify/<id> without asking us. Next: a specialty-matched reviewer, named on the record; today: the sandbox runs synthetic samples in your browser.
Sign and verify a demonstration record →What does it cost?
Nothing is for sale on this site, and there is no checkout or subscription. The sandbox has nothing to buy. Production terms are set in the pilot and follow a BAA.
How companies work with ClinicalSwipe →Can I send real patient data?
Not yet. The sandbox accepts synthetic or de-identified outputs only. A Business Associate Agreement is offered to qualifying covered entities; production with patient data starts after one is.
The compliance matrix, stated honestly →How fast is a review?
Each review is designed for three to five minutes of physician time. There is no production volume yet, so we publish no measured turnaround and promise none in the sandbox.
How a review works, in 90 seconds →Request a pilot.
Five fields. The founder reads every one. If the answer is not yet, you will be told what has to be true first.