The Problem with AI in Healthcare Today
Healthcare was always the sector where getting the incentives wrong would cost lives. For decades, the dominant business model — fee-for-service, upcoding, billing complexity as a competitive moat — has steadily increased the share of every healthcare dollar that goes to administration rather than actual care. Today, the administrative overhead of American healthcare consumes roughly 34 cents of every dollar spent.
Artificial intelligence is being deployed across this system at scale. The question is not whether AI will transform healthcare — it already is. The question is: who does it serve? In most current deployments, the answer is not the patient.
AI is being used to find more things to bill for, to automate denials of care, and to maximize risk-adjustment revenue for health plans — all at scale that was previously impossible. At the same time, the tools that could actually reduce costs and improve outcomes — accurate documentation, earlier diagnosis, reduced redundant testing — are being subordinated to billing optimization.
AI in healthcare must be evaluated by a single standard: does it reduce the cost of care and improve outcomes for patients, or does it increase administrative revenue for payers and providers? Current deployment overwhelmingly serves the latter. That is a policy failure, not a technology failure — and it is correctable.
Where AI Is Being Used — and How It's Going Wrong
1. Ambient Listening & Upcoding
Ambient listening devices are now deployed in clinical settings at major health systems. These devices record and transcribe patient-provider conversations, then use AI to identify billable diagnosis codes the provider may not have explicitly documented. The stated benefit is reducing administrative burden on clinicians. The actual effect, in many deployments, is systematic upcoding — identifying the highest-paying code supported by any phrase in the conversation.
Patients have no meaningful notice that they are being recorded for billing optimization. The recordings and transcriptions exist in a regulatory gray zone. No federal law currently requires disclosure to patients that ambient AI is active during their visit, or that their conversation is being used to determine billing codes.
2. AI-Driven Denial of Care
Health plans are deploying AI to automate prior authorization denials at a scale no human workforce could achieve. In documented cases, AI systems have denied claims at rates far exceeding the historical denial rates of the plans that deployed them — with internal audits and court filings revealing that the AI was configured with denial rate targets rather than clinical appropriateness standards.
When a human reviews a denial, they can be held accountable. When an AI denies care, accountability disappears. The patient receives a letter. The plan keeps the premium.
3. Risk Adjustment Manipulation
Medicare Advantage pays health plans a per-member monthly capitation rate adjusted for patient risk — sicker patients generate higher payments. AI is now deployed at scale by health plans and their vendors to mine medical records for conditions that increase risk scores, regardless of whether those conditions are being actively managed.
When implemented responsibly, this improves care: accurate documentation means patients get the right interventions. When implemented aggressively, it is fraud: fishing for codes that inflate revenue without improving care. The line between the two is a matter of governance and design — and most current AI deployments in this space have weak or no guardrails.
4. The Interoperability Shell Game
Federal mandates require health systems to support interoperable data exchange (FHIR/USCDI standards). In practice, dominant EHR vendors have achieved bare-minimum compliance — allowing patient access to data while blocking the two-way data exchange that would enable genuine care coordination and competitive innovation.
The result: AI companies that could improve care can't access the data they need, while the companies already embedded in major health systems face no pressure to improve. Interoperability mandates without enforcement are not mandates — they are a competitive shield for incumbents.
The PTI Policy Framework
PTI is developing model legislation for state and federal application. Our framework is grounded in a principle the healthcare industry has never fully accepted but can no longer avoid: when AI makes a decision that affects a patient, there must be a human accountable for that decision.
The Cost Curve Argument
Every discussion of AI in healthcare eventually returns to a promise: that AI will bend the cost curve — reducing the administrative overhead that consumes a third of every healthcare dollar. This argument has been used to justify regulatory restraint and to resist transparency requirements. PTI believes it is, in its current form, a false promise.
The cost curve bends when the dollar saved on administration reaches the patient as lower premiums, better access, or improved care. In the current model, the dollar saved on administration becomes margin for the health plan or revenue for the vendor. The patient sees none of it.
For Texas specifically, this is not an abstract concern. The state is one of the largest Medicaid payers in the country. Every percentage point of administrative overhead in the Medicaid program is a dollar not spent on a road, a school, or a child's healthcare. The state has both the standing and the obligation to set standards for how AI is used in the programs it funds.
Texas has an opportunity to establish a national model. HB 1188 and SB 815 were important first steps — establishing the principle that a human must be in the loop. The next session should go further: defining what that human review must actually look like, who is accountable when it fails, and how the state measures whether AI is actually reducing costs for patients or simply extracting more revenue from the Medicaid system.
A Summary Framework
| Issue | Current State | PTI Recommendation |
|---|---|---|
| AI clinical documentation | Often entered without human sign-off | Mandatory licensed clinician approval before record entry |
| AI-driven care denials | Fully automated in many plans | Require licensed clinician review for all adverse decisions |
| Ambient listening | Unregulated; minimal or no patient disclosure | Explicit consent required; billing use must be disclosed |
| Risk adjustment AI | Widely deployed; audit standards weak | Multi-source evidence requirements; annual independent audits |
| EHR interoperability | Bare-minimum FHIR compliance; one-way access | Bidirectional exchange mandated; penalties scaled to market share |
| AI spending transparency | Not reported | Annual public disclosure for Medicaid contractors and large plans |
Get Involved
PTI is actively seeking collaboration with legislators, healthcare policy experts, clinicians, health plan executives, and patient advocates to develop and refine this framework. If you are working in this space — or if you have been harmed by the practices described here — we want to hear from you.
For legislative briefings, policy partnership inquiries, or to contribute to our healthcare AI research, contact [email protected].