Price Protection Is the First AI Feature to Buy

Many of the AI vendors a health system buys from pay the model companies underneath them, so they do not control their own costs either, and by one research firm's account those model companies are not yet profitable at today's prices. Systems that find the price each AI business case can bear, and ask vendors for fair price protection, take on less risk.

Environment

Most of the AI tools a health system buys run on a small number of foundation models, and many of the vendors selling those tools pay the model companies by usage. Ty Aderhold, an AI expert at Advisory Board, said on Radio Advisory in October that the independent companies behind those models, such as OpenAI and Anthropic, "are not making money" at the prices they charge today and depend on investors to keep operating, though cheaper computing could lower their costs. His is an outside view from a research firm that sells to health systems. He expects these companies to follow the usual path of investor-backed technology firms toward profit, through "either ad revenue or increasing prices," and pointed to Uber, whose fares stayed low for years and rose once it no longer relied on investors (Radio Advisory #318).

He also described how that plays out inside a health system. Clinicians who have rebuilt their work around ambient listening will want to keep it, so if its price doubled in two years, a scenario he offered, the system would face a hard choice about taking it away. He reports a common concern in the market: token-based pricing makes the monthly bill hard to predict when a rollout grows from 10 users to 1,000.

Leaders closer to the work share the concern. Dr. Bradley Locke, chief medical information officer at Prevea Health, called AI the biggest headwind he is watching and said how to pay for it is one of the problems his team wrestles with (Becker's Healthcare Podcast).

Impact

An AI business case built on today's price is a forecast about someone else's finances, often two companies removed. Many of the AI vendors a health system buys from pay the model companies underneath them, so they do not control their own costs either, and a system that signs without price protection carries that forecast risk itself. The return in most business cases comes from avoided hiring, saved clinician time, or fewer denials, each measured against the vendor's price at signing. As staff and workflows adapt to a tool, the cost of switching away rises, and the vendor's bargaining position improves with it.

Two kinds of exposure follow. The first is a price the vendor raises at renewal, after the system has built its operations around the tool. The second is usage that patients, staff, or connected systems drive, billed per call or per token at a rate agreed before anyone knew how much traffic would arrive, which shows up as an unbudgeted variance months after the operating budget was set. These are operating costs, so every increase lands on that year's margin in full. Aderhold's doubling scenario sets the low end of the stress test, and the three-times margin is my own, because usage can grow from 10 users to 1,000.

Imperative

CFOs should test each AI business case at two to three times today's price and ask vendors what share of their price is model cost.

Insight

Every business case rests on conditions that have to be true for the return to hold. Most AI business cases test adoption and leave price untested, though price is the one condition the system can control through the contract.

The price protection a vendor will sign tells the system how well that vendor knows its own costs. No vendor controls what the model companies charge, so a vendor can reasonably refuse to cap a cost it cannot set. A vendor that can say what share of its price is model cost, cap its own fee, and pass the rest through at a stated formula has measured its risk and priced its margin, and one that cannot answer is handing an unmeasured risk to the system at renewal. That makes price protection a reason to choose one vendor over another, scored in the same column as clinical results.

Initiative

Start with the AI pitches already in front of a committee. Before a business case goes to approval, finance reruns the return at two and three times the quoted price and at half the promised realization. The goal is each case's break-even price, the multiple of today's price at which the return disappears, weighed against how hard the tool would be to replace. The riskiest tools break even close to today's price and sit inside the daily work of hundreds of clinicians. In the same review, ask each vendor what share of its price is model cost, as a range if it will not give the figure, and whose contract that capacity is bought under. A case that pays only at today's price needs price protection before the system signs, and the others can move ahead on standard terms.

For the tools that need it, ask for a multi-year price cap. In practice this is a renewal option the system holds for several terms, with each increase on the vendor's own fee capped at a stated percentage and model costs passed through at the agreed formula. A cap worth having defines the unit the system uses, such as a finished note or a coded chart, covers successor products with the same function, sets usage bands, and limits per-call billing, and the same formula passes falling model prices on to the system. Pair it with exit terms for data return and portability, because a cap is worth little once a vendor fails or an acquirer retires the product. Ask hardest on the tools that clinicians and departments will build their work around, such as ambient listening and coding assistants, because the cost of walking away from them grows each year. In return the vendor gains a longer commitment, committed volume, and revenue it can forecast, which makes a cap on its own fee easier to sign. Put the same terms on the RFP scorecard and weight them alongside clinical value and integration.

Systems with an AWS or Microsoft cloud commitment have one more lever. Some vendors can run their AI in the system's own cloud account, so the model usage is billed under the system's cloud agreement and counts toward a commitment the system already owes. I have seen vendors set this up with health systems already, and it is a fair question to ask any AI vendor in a pitch. Eligible purchases through Microsoft's marketplace also count toward an Azure commitment (Microsoft), and AWS offers a similar benefit on eligible listings under the terms of each agreement, so check yours. The model price then sits on the system's own bill, where its cloud agreement sets the terms, and the vendor's cap covers its software fee. The setup needs a security review, a business associate agreement that covers the model service, and someone at the system who owns model updates.

Worth Repeating

Many of the AI vendors we buy from pay the model companies underneath them, so they do not control their own costs either. Before we sign, we should know the price each AI tool can bear and ask for protection on the ones we could not easily give up.

Questions for your leadership team

1. If your largest AI contract doubled in price at renewal, which workflows could you give up, and which could you not?
2. Which of your AI contracts bill you for usage that patients, staff, or connected systems drive, and who would see a spike first?
3. Which AI pitch in front of you today breaks even closest to its quoted price, and what share of that vendor's price is model cost?

Michael Alden — Managing Partner, Nexcurve