RatedWithAI

RatedWithAI

Accessibility scanner

Free scan →
AI Legal & ComplianceAugust 7, 2026

The AI Was Free. That Is the Problem.

Health care fraud and abuse law is indifferent to how novel the technology is. It asks two questions — is something of value moving, and is the recipient in a position to send business — and an AI go-to-market motion answers yes to both in ways a SaaS playbook never had to think about.

Anti-Kickback Statute

Intent-based criminal statute. Prohibits knowingly and wilfully offering, paying, soliciting or receiving remuneration to induce or reward referrals of items or services payable by a federal health care programme. Safe harbours are voluntary — missing one is not automatically a violation, but it removes your certainty.

Physician Self-Referral (Stark)

Strict liability. If a physician has a financial relationship with an entity, referrals for designated health services to that entity are prohibited unless an exception is satisfied in full. Intent is irrelevant, and substantially complying with an exception is not complying with it.

What follows is four commercial structures that AI health vendors and health systems actually build, taken apart the way outside counsel takes them apart: where the value is, where the referral is, and what would have to be true for the arrangement to be defensible.

Four Arrangements, Taken Apart

The donated deployment

High

A vendor gives a referring practice free access to an AI imaging or triage tool, framed as adoption seeding.

Where the remuneration is
The software itself, plus implementation, integration, support and training — all of which have market value the recipient would otherwise pay for.
Where the referral is
The practice sends studies, procedures or patients to the vendor's affiliated facility, or to the health system funding the donation.
What would make it defensible
Fitting a recognised exception in full rather than approximately: documented fair market value, a recipient contribution where required, recipients selected on criteria unrelated to referral volume, and a written agreement predating deployment. The common failure is a pilot that started before any of it was papered.

Usage pricing that tracks reimbursable volume

High

Per-scan, per-encounter or per-enrolled-patient pricing, described as consumption-based and therefore fair.

Where the remuneration is
The payment itself, flowing in whichever direction the deal runs. Where the vendor pays the practice, the exposure is direct.
Where the referral is
The unit of consumption is the same unit that generates a federally reimbursable claim, so the payment moves with referral volume by construction.
What would make it defensible
Compensation set in advance, at fair market value, not varying with the volume or value of referrals of federally reimbursable business. Genuine consumption pricing for a tool the practice buys for its own use is a different structure from a success fee tied to what gets ordered — and the paperwork should make which one it is obvious on its face.

Vendor-funded care navigation and outreach

High

The vendor funds an AI outreach or navigation layer that identifies eligible patients and books them, at no cost to the practice.

Where the remuneration is
Free staffing-equivalent labour. Work the practice would otherwise pay for is being performed at the vendor's expense.
Where the referral is
The navigation logic directs patients toward specific services, sites of care or products, and the funder benefits from where they land.
What would make it defensible
Neutral routing the funder cannot influence, patient-choice language that is real rather than nominal, no per-patient success payment, and independent clinical criteria for who is contacted. Beneficiary inducement rules are also live here, since the patient is receiving something of value too.

Sponsored logic inside clinical decision support

Severe

A recommendation, order set, formulary suggestion or preferred-provider ranking inside the clinical workflow, partly funded by a party that benefits from the recommendation.

Where the remuneration is
Payment for placement or for influence over the recommendation, however it is invoiced — development funding, data licensing, co-marketing.
Where the referral is
The recommendation surface is the referral channel. It reaches the ordering decision at the moment it is made.
What would make it defensible
Very little, if the influence is paid for. The workable structures separate funding from ranking entirely: no sponsor input into logic, clinical criteria published and reviewable, and any sponsorship disclosed at the point of display rather than in a settings page.

Why Software Vendors Get This Wrong

Every instinct a good SaaS operator has is a liability here. Free trials, generous pilots, usage-based pricing, referral incentives, co-marketing funds, land-and-expand seeding — the entire modern playbook consists of moving value toward the party who decides what gets bought and ordered. In most industries that is called go-to-market. In this one it is the elements of an offence.

Three specific translation errors recur:

  • Treating "no cash changed hands" as safety. Remuneration is anything of value. Engineering time, free integration, waived implementation fees, conference sponsorship, data licensing at above-market rates and advisory board seats are all value.
  • Assuming a safe harbour covers the whole product. Protections are drawn narrowly around defined categories and conditions. A bundle that contains one protected component and three unprotected ones is not a protected bundle.
  • Papering it after the pilot. Most exceptions require a written agreement, signed, covering the identified items, for a defined term, with compensation set in advance. A retroactive agreement cannot make compensation set in advance.

The AI-Specific Wrinkle: Influence You Cannot See in the Contract

Sponsored placement in clinical software is a known enforcement theme, and the historical cases involved rules — an alert that fired for a particular product because someone paid for it to fire. A model changes where that influence can hide.

A ranking produced by a learned function can be shifted by the choice of training corpus, by which outcome the objective optimises, by which providers are represented in the feature set, and by a retrieval index that happens to contain more material from one party. None of those appear as a line in the recommendation logic, and none are visible in the contract that funded the work. That makes governance documentation — who contributed data, who influenced the objective, who reviewed the output distribution — the only place the answer lives.

The practical test that survives this: if a party who benefits from a recommendation contributed money, data or engineering to the system producing it, you need an independent, documented reason to believe the recommendation is unaffected. Absence of a rule is not that reason.

The Three Questions Counsel Asks

1

What is moving, and what is it worth?

Enumerate every item of value in both directions, including the ones nobody invoices: implementation labour, integration engineering, training, data, support, marketing, board seats, equity. If it would appear on a budget were the other party paying for it, it is remuneration.

2

Is the recipient in a position to generate federally reimbursable business?

This is a factual question about the recipient's role, not about your intent. A physician group, a hospital, a lab, an imaging centre, a pharmacy, a device supplier and increasingly a care-navigation vendor can all be sources of referrals. If the answer is yes, the arrangement needs to fit somewhere.

3

Does the compensation vary with volume or value of that business?

If yes, the structure needs to change rather than be documented more carefully. Set-in-advance, fair-market-value compensation that does not take referral volume or value into account is the shape almost every relevant exception and safe harbour requires, and it is a design constraint, not a drafting one.

What Enforcement Actually Looks Like

The exposure is rarely a criminal prosecution of a software company. It is a False Claims Act case, frequently brought by a relator, on the theory that claims tainted by a prohibited arrangement were false claims — with the vendor drawn in as having caused their submission. That path is why the amounts involved are disproportionate to the value of the software, and why the person who files it is often a former employee who watched the pilot go live without paperwork.

It is also why the cheapest control is cultural rather than legal: a sales organisation that knows a free pilot with a referral source requires a signed agreement first will generate far fewer of these than a compliance memo nobody in revenue has read.

Related Reading

Check What Your Site Says About Your Deal Structure

Health AI marketing pages promise things the compliance team would not sign off on — free for referring practices, we only get paid when patients convert, preferred network placement, no cost to your clinic. Those sentences are read as evidence.

See every pricing, partnership and compliance claim your site makes in one pass. Run a free scan and check each against the arrangement you actually signed.

This article is general information and not legal advice. Fraud and abuse analysis is intensely fact-specific, safe harbour and exception requirements are detailed and conjunctive, and state law adds further prohibitions that apply regardless of payer. Engage qualified health care regulatory counsel before structuring or relying on any arrangement described here.