AI implementations cost 3 to 5 times the advertised subscription price. In a fixed-reimbursement market, that cost has nowhere to go.
The market is excited about AI in healthcare. The unit economics tell a different story.
Start With a Simple Number
Take a best-case scenario for a clinical service. Assume the unit cost is $100 per patient encounter. The provider's operating margin is $10 — ten percent. In healthcare, where industry-wide median operating margins historically hover closer to 2 to 3 percent, a 10 percent margin is a healthy baseline.
Now introduce AI into that workflow.
The vendor quotes a license fee. That fee is the smallest part of what you will actually pay. Enterprise pricing benchmarks show that AI implementations cost 3 to 5 times the advertised subscription price once integration, data pipeline management, ongoing model retraining, compliance, and human-in-the-loop validation are included (AI Software Cost: 2025 Enterprise Pricing Benchmarks for Manufacturing Leaders). In one anonymized institutional deployment, 63 percent of total AI expenses came from data pipelines and GPU cluster management — none of it captured in the vendor proposal (The Real Cost of AI: Calculating the Total Cost of Ownership for AI/ML Systems).
On a $100 unit of care with a $10 margin, the fully-loaded AI cost on a per-encounter basis — once that 3 to 5x multiple is applied to what a comparable legacy system costs to run — lands conservatively between $10 and $15. The margin goes to zero or below.
The Constraint That Doesn't Go Away
Healthcare pricing is externally set. Reimbursement rates are negotiated with payers or determined by CMS. A provider cannot pass an AI surcharge to patients, and payers do not automatically reimburse for technology investments. The Bipartisan Policy Center has documented that fragmented coverage pathways and limited benefit category alignment create significant uncertainty for both providers and AI developers — with most clinical AI tools currently relegated to temporary Category III CPT codes without guaranteed reimbursement (Paying for AI in U.S. Health Care, Bipartisan Policy Center, 2026).
There is no revenue-side release valve for this margin pressure.
AI must fully displace an existing cost layer — not augment it, not run alongside it. The legacy system has to be decommissioned.
The throughput case is real: if AI allows one clinician to handle ten patient encounters per hour instead of one, with high accuracy and minimal human review, fixed costs amortize across more volume. But that only works if the existing platform is eliminated, if accuracy is genuinely high enough to reduce human oversight, and if physical capacity — beds, OR time, treatment slots — exists to absorb the additional volume.
In most deployments, none of those conditions fully hold. The EHR stays. The legacy workflow stays. The staff stays. AI becomes an additional cost line, not a substitution. Margin shrinks.
What the Full Cost Stack Looks Like
The license fee covers less than half of total AI spend (Deloitte Emerging Technology Trends). Two categories drive most of the variance: upfront integration costs and recurring operational expenses that compound annually.
According to Deloitte's Emerging Technology Trends research, most enterprise budgets underestimate true AI total cost of ownership by 40 to 60 percent. This is where that gap lives.
This cost structure exists across all industries. In industries with thin margins and constrained pricing, it compounds into a structural problem. Importantly, the vendor does not change the math — whether AI is introduced by a startup or an incumbent platform already embedded in the workflow, the underlying reimbursement environment does not flex to accommodate it.
The Investment Implication
Clinical AI platforms are attracting significant capital based on technical sophistication and clinical outcomes data. Both are real. What is less clear is whether clinical value and commercial sustainability are the same thing.
A platform targeting oncology or specialty care is selling into a buyer base operating on 1 to 3 percent margins (2025: The State of AI in Healthcare, Menlo Ventures), with no ability to raise prices and limited appetite to fully decommission legacy systems. The platform either prices below its cost to stay affordable, or prices for value and slows adoption.
From a capital allocation standpoint, durable margin expansion is structurally blocked on both sides: the cost of delivering AI is high, and the ability of buyers to fund it is constrained. Growth projections that do not account for this tend to be optimistic.
The healthcare AI businesses with the clearest economics are those targeting specific administrative or operational processes — high-cost, well-defined, with an existing budget line that AI directly replaces and eliminates.
One Question Worth Asking
Before approving an AI investment: what specific existing cost does this replace, and will that system actually be decommissioned?
If the answer is no, or not yet, the economics do not close.
This is why the clearest near-term opportunity in healthcare AI is not clinical — it is administrative. Processes such as outsourced medical billing, manual prior authorization, and transcription services carry defined budget lines, no liability risk from replacement, and no physical capacity constraints. These are the workflows where an agentic AI system can sever an existing cost cleanly. That is where the unit economics actually work.