Health systems are already scrambling to rein in AI-based token costs, and the proliferation of AI agents is poised to further complicate that work.
Jeff Gautney, senior vice president and CIO of Rush University System for Health in Chicago, spoke to this challenge during a Sept. 16 panel at Becker’s 11th Annual Health IT + Digital Health + RCM Conference.
Rush is more than 18 months into its AI agent journey. The health system is building patient-facing agents on Salesforce’s Agentforce platform to answer routine website questions such as clinic hours and locations and to handle prescription refills, the second-most common reason patients contact its support line.
While the goal is to free up call center staff for more complex cases, Mr. Gautney has warned that AI agents could end up costing more than the human labor they are replacing if health systems do not exercise strong financial operations discipline.
Nearly every core system a health system runs now ships its own agent development tool, he said, and the pricing models are moving from subscriptions to token consumption. That means the bill is set by how often agents run and how much they process — variables that aren’t fixed when the contract is signed. It also means there is no easy way to see total AI agent costs because each vendor reports usage in its own administrative console, leaving executives to piece together totals from disparate systems.
“There’s not a common way across those platforms to really do financial modeling,” Mr. Gautney said.
What the industry needs, in his view, is the discipline that health systems built to control cloud spending, applied to agents. Cloud FinOps teams track consumption in near real time, forecast it and charge it back to the department that incurred it. Agents would need one thing more: a way to model what one will cost before it is switched on. Some vendors are beginning to move in that direction. In the meantime, Mr. Gautney sees the cost exposure sitting with the people authorized to create agents.
“Training the folks who have access to turn those agents on, and having a financial operations way of modeling it, much like we have for cloud services, is a critical next step,” he said.
At Rush, that training is tiered: a 100 level required of all IT staff, a 200 level for the citizen developers given access to governed datasets and agent-building tools, and a 300 level for employees working in AI systems daily.
Some of that training comes down to a single decision: which model to point at a given task.
“Everybody wants to use the frontier model because they just assume that it’s the best. But it’s also orders of magnitude the most expensive way to do something,” Mr. Gautney said. Some tasks, he added, do not require an agent at all. A simpler model, or an ordinary search, would suffice.
Another cost nobody has priced arrives well after the agent is built. An agent is not a subscription that lapses or a license reclaimed at termination. It is working software, written by someone who may not be at the organization in a year, running against systems that will change underneath it. Accountability for it, Mr. Gautney said, cannot sit with IT alone.
“Somebody in supply chain develops an agent, they leave the organization and move on. But this agent is running. When the agent breaks, who does the help desk call to get support?” he said.
He expects that to fragment into hundreds or thousands of agents across a health system, each with its own dependencies and each needing some form of support model, few of them built by anyone in IT.
“This is going to force us within the IT industry to rethink our own organizational models and how we engage with our customers, because our customers are going to have tools that typically have only been our domain,” Mr. Gautney said. “It’s exciting, but it is going to really tax us to think differently.”
