The Procurement Trap

A 540-bed health system signs a 3-year contract for an AI operator for sepsis prediction after completion of an 8-week pilot program at one of their hospitals. The vendor's board materials listed a 30% reduction in mortality at four other sites where the product was not yet operational (they were probably omitted for model drift), which excited the board for the contract's approval. The deal employs standard healthcare IT contract language. The contract defines an 'EHR change' to be a code modification or a UI adjustment, and a model is treated similarly. 18 months later, after the tool was deployed across 4 hospitals in the health system, model drift changed the mortality reduction from a predicted value of 0% to -6% and still declining. The board is approving the deal, as they didn't have alternatives to offer.

The decision made by the board to procure the tool 22 months ago is the decision the health system is obligated to accept for the next three years.

Pilots are by design meant to succeed. The vendor designs the pilot to their liking, controlling all facets from cadence of implementation to the scope of integration, the refresh of training data, and the criteria to declare the pilot a success. Reference calls are made to early adopters whose deployments are too new to have begun to experience the concept of drift, and pilot datasets have been pre-cleaned. The data and the clinical workflow modifications are decoupled. The workflow changes do not occur in the pilot, and they do not occur in the timeframe of the modeling exercise. Pilots are kept short in duration in order to avoid the conditions under which drift occurs.

What the pilot does not disclose is how the local patient population is served in the long run. One of the most prominent critiques in the Coalition for Health AI Blueprint, Draft 1.0, is poorly structured pilots which make a sharp contrast between what is observed during the pilot and what is expected in day to day operations. The AHRQ patient safety reports cite the absence of a continual post-deployment assessment, and say that the contract stipulations should include both monitoring and evaluation. Most health care organizations end up signing the pilot contracts, which limit the monitoring and evaluation only to the pilot phase.

Reference calls are also unable to fill the gap. Vendors build reference lists with systems in the early stages of deployment. Systems in this stage are less likely to have gaps. By the time the calls are made, the system that is most likely to identify procurement gaps is the one that is most mature, and is likely the one that is not on the reference list. For a nascent AI procurement team, it is normal to call other vendors that have deployed their first AI systems to do the due diligence. It is certain that both vendors are not in the renewal cycle.

The standard language used in healthcare IT contracts for EHR and similar systems, which typically use a one-time configuration that runs, has been adapted to contracts for AI. Contract language for AI models must explicitly state and codify certain provisions. This includes defining who decides the timing of 'sufficient' changes to justify an update, the funding source for such changes, and the performance evidence that will be provided by the vendor, as well as the means by which performance will be monitored. The cessation of rectification rights must be explicitly contracted. Binding performance related Service Level Agreements must endure the renewal cycle.

The ONC HTI-1 Final Rule mandates that predictive decision support systems disclose their training data sources. This provides buyers with a way to ask better questions before buying. However, the disclosure alone does not address the issues with contracts. Where the HTI-1 disclosure has been incorporated into evaluation frameworks, the training data questions vendors were previously silent on are now required to be answered, and the provided answers are expected to result in changes to the contractual arrangements that will follow. Most evaluation frameworks developed by procurement teams do not yet include the HTI-1 disclosures.

There is a separate but related issue regarding the vendors holding all the IP for any improvements made to the training. The vendor owns any improvements made to the model using a health system's data, and other customers have to pay for that model improvement. The customer whose data was used to improve the model does not benefit. These improvements are specifically related to the AMA's augmented intelligence policy. Concerns regarding the IP issue and AI vendor oversight are noteworthy, but as of 2026, the specific resolution of these contracts has not been addressed by policy.

Renewals occur anytime from 18 to 24 months after initial contract signing. Model drift is now at a level that is acceptable to the vendor and not to the health system. Costs of integration are now significant, as clinical teams have been trained to work with the model perform outputs. The original contract offers no pricing for the costs associated with undoing everything.

Vendors are familiar with the numbers. Finalized renewal rates come with a notice that operational costs have been updated or a higher tier includes greater levels of functionality. The health system can either renew on the vendor's terms or absorb the costs of a hastily procured replacement Integration Write Off. Most systems renew.

ECRI Institute considers the use of AI in health care as the number one risk and differential for 2025. The accompanying explanation in the report mentions gaps in Lifecycle management, including contract structures and renewal practices, as the primary driver of the risk in question. The trend is identical across systems, regardless of bed count and geography. Standardized contract templates are the main culprits. The Robert J. Margolis Institute's work on the composition of AI committees identifies managing vendor relationships as one of the differentiating factors of committees that survive beyond their first year.

The terms of reference end the conversation. After eighteen months, the board will have limited options. The contract terms are owned by the CAIO, as discussed in Edition AP, along with legal counsel, the CFO, the CISO, and the procurement office. The five readiness questions from Edition AQ are transformed into five procurement questions: What is the contract's halt authority? Where does local validation align in the assessment clauses? Who will own IP enhancements in three years? What will the exit costs be at months six, twelve, twenty-four, and thirty-six? What is the structure of renewal pricing at the time of signing?

The AHA Trustees Toolkit on AI in health care recommends vendor selection as one of the 2024 board agenda items but, to date, most procurement teams have not included these questions in their standard evaluation rubric. In the Coalition for Health AI, Blueprint v1.0, procurement gaps are addressed using a shared evaluation methodology with a multi-system, pre-purchase evaluation, designed as a counterbalance to vendor-controlled reference calls. Vendors have promised a 30% reduction in sepsis mortality, and the promise is validated with each iteration of the pilot. The renewal arrives in 18 months with new pricing and unanticipated integration costs. If the pilot begins with procurement decisions, a transformational change occurs with a CAIO's presence, a readiness diagnostic completed, the procurement office aligned on AI evaluation, and an optimized contract. The board accepts the renewal as it is the least bad option. The changes in procurement for 2026 shape how we understand the alternative in 2029.

This edition draws on the Coalition for Health AI Blueprint (Draft 1.0 and v1.0) on pilot evaluation and shared evaluation methodology, AHRQ patient safety reports on post-deployment assessment, the ONC HTI-1 Final Rule on training data disclosure for predictive decision support systems, the AMA augmented intelligence policy on vendor AI oversight, the ECRI Institute Top 10 Health Technology Hazards report for 2025, the Robert J. Margolis Institute for Health Policy at Duke University on AI committee composition, and the AHA Trustees Toolkit on AI in health care on vendor selection. Related editions: Issue 43 (The Chief AI Officer Mandate) and Issue 44 (The AI Readiness Diagnostic).