Culture
HR Series - Building competitive advantages with Alice Yoo Le Clair
What responsibility, if any, do HR leaders have for community building within their organization;...
Read more >
Assign ownership and review AI spending
Usage-based AI services give health systems a spending variable that can change with daily work. The bill depends on what teams ask the service to process, which models they use, and how often a workflow calls them. Costs generated within one function may still appear on a central technology budget.
This is a familiar management problem in a new application. Procurement establishes the rates and terms, but operational choices determine how much the organization consumes. A pilot budget becomes less informative as access expands and workflows change.
“Procurement establishes the rates and terms, but operational choices determine how much the organization consumes.”
Assign an owner who can connect usage to the work it supports. Give that owner a budget, cost information, and a regular review with finance and the teams making technical decisions.
Know what contributes to the bill
Model application programming interfaces (APIs) commonly charge by token, a unit of text or other data the model processes. Input and output have different rates, and the choice of model can materially change the cost. Anthropic and OpenAI publish separate rates for these categories. Tool use, storage, infrastructure, and other services can add charges beyond the model request itself.
Two options are worth examining for eligible work. Prompt caching lets a service reuse a previously processed portion of a request, such as policy material sent repeatedly. Cache reads can cost substantially less than ordinary input, although savings depend on reuse and the provider's terms; Anthropic also charges for cache writes. Batch processing offers a 50 percent discount on eligible API processing when an immediate response is unnecessary.
These options can reduce processing cost while retaining the chosen model and task. They do not halve every component of the implementation budget. Ask technical and finance teams to review the same workload, including how often the cache is used and which jobs can tolerate a delayed response.
Match the model to tested task requirements
A lower-cost model may meet the requirements of a particular task. Establish that through evaluation before making it the default. Even a short summary can carry clinical or operational consequences if it omits important information.
Define the accuracy, reliability, and review requirements for the use case. Compare models against representative examples, then account for retries, corrections, and human checking. A low price per request is useful only if the resulting work meets the required standard at an acceptable total cost.
Where several models are available, specify which tasks each may handle and when work must move to a more capable model or a human reviewer. Department-specific assistants can make the permitted scope clearer, but their cost still depends on the model, information retrieval, maintenance, and volume of use.
Budget for changes in the workload
Lower processing costs do not guarantee a lower invoice. Gartner's March 2026 forecast anticipates a substantial decline in provider-side inference cost by 2030, while warning that more demanding applications will consume more tokens. Its August forecast separately projects rising inference cost per agentic workflow through 2028 as systems undertake more complex, multistep work.
“Lower processing costs do not guarantee a lower invoice.”
Those are forecasts about technology economics, not guaranteed reductions in a health system's contracted prices. An expansion proposal should estimate the number of model calls per completed task, the amount processed in each call, the model mix, and the likely need for retries or additional tools.
Check whether the organization can attribute the resulting spend. In Flexera's 2026 research, 31 percent of respondents reported accurate visibility into AI software, and 59 percent said wasted AI spending had increased. Those cross-industry findings provide a reason to examine local visibility, rather than an estimate of the health system's own waste.
Use alerts and limits together
A spending alert needs a named recipient, a response time, and a decision the recipient is authorized to make. Without those, it can document an overrun without changing it. Hard limits can also be useful, provided the organization accounts for service continuity and defines how exceptions are handled.
When consumption rises, investigate the work behind it. A temporary increase supporting a time-sensitive implementation is different from repeated failed requests or use of an unnecessarily expensive model. Retain usage records for blocked requests as well as completed ones so a limit does not end the investigation.
Build the review around completed work. What did the service help produce? Did the result meet the required standard? What did it cost, including review and correction, compared with the previous process? Those answers give leaders a basis for deciding where to expand, redesign, or limit use.
Connex convenes health system finance and operations executives in Trusted Peer Groups. Ask peers who owns AI consumption, how they allocate the cost, and what they learned from their first unexpected bill.
See how peer CFOs and COOs are approaching this.
