Invest in outcomes, not usage: why token efficiency matters in healthcare AI
When ECU Health started budgeting for its first AI agents, it had no reliable way to tell whether a given month of usage would cost $100 or $1 million. That’s a problem more healthcare organizations are about to face.
ECU Health’s experience isn’t an isolated one. Recent benchmarking from Ramp, which tracks AI vendor payments across thousands of businesses, put tangible numbers on a problem most finance teams have felt but couldn’t yet quantify:
Median monthly AI token spend (April 2026): $2,246
Average monthly AI token spend (April 2026): $140,842
That gap exists because a small number of heavy users, often a single team or an automated workflow running unchecked, pull the average far above typical usage.
Token usage across tracked businesses grew more than 1,000% between January 2025 and April 2026. Spend grew more slowly, since per-token prices fell over the same period, but it still climbed nearly 500%.
Healthcare has more at stake in getting this right than most industries. Payers and health systems aren’t running a handful of internal AI workflows, they’re potentially running AI conversations across member populations in the hundreds of thousands or millions, every day, indefinitely. At that scale, it becomes the difference between a program that’s sustainable and one that stops being affordable the moment it succeeds.
The two questions every AI buyer is actually asking
Bret Taylor, OpenAI chairman and Sierra co-founder, put it plainly in a recent CNBC interview: in conversations with roughly 100 CEOs each month, two questions come up constantly:
“Tokenomics.” Is a company actually getting value for its token spend?
Sovereignty. What durable competitive advantage is a company building as the underlying models keep improving?
On the tokenomics question, Taylor pushed back on a common assumption: cheaper, open-weight models aren’t automatically the answer, because they can be less token-efficient than frontier models for a given task, which erases the savings a lower price tag seems to promise. Rather than a model that’s the cheapest per token, the real value comes from how many tokens, and how much reasoning, a task actually requires. As Taylor joked in the same conversation, running every question through the most expensive model available is a bit like driving a Ferrari to the grocery store.
Taylor’s own company reflects that thinking: Sierra prices its AI agents by outcome, a completed loan origination, a completed preauthorization, rather than by the token. That mirrors a shift Taylor argues the broader industry is heading toward: away from metering AI the way early cloud computing metered raw CPU cycles, and toward paying for the result the AI actually produced.
Why costs spiral
Ramp’s research points to a specific, avoidable pattern behind rising bills: teams default to the most capable model available for every task, including the routine ones that don’t need that much reasoning power.
Premium, high-cost models made up just 45.8% of tokens consumed across the businesses Ramp tracks in April 2026, but 55.9% of total spend, and that premium-model cost share had risen from under 6% less than a year earlier. Model tier migration, upgrading to a more expensive model without re-examining whether the task actually requires it, was identified as the single biggest driver of unexpected cost increases, sometimes multiplying cost by 10–100 times for the same workload.
The fix isn’t complicated:
Test before defaulting to premium. Running routine use cases on a lighter, cheaper model first typically preserves quality while cutting cost by an order of magnitude.
Cache what repeats. Caching repeated prompts instead of reprocessing them from scratch can cut costs by roughly five times for identical work.
What healthcare leaders are learning in practice
This isn’t a hypothetical problem for healthcare, and ECU Health isn’t the only one experiencing it.
Advocate Health, building agents on the same Epic Agent Factory platform as ECU Health, is running into the oversight side of that same challenge. Advocate holds a strict human-in-the-loop rule for anything its AI Studio creates: according to the health system’s chief data and AI officer, AI is never permitted to place an order or make a diagnosis on its own.
That rule holds regardless of case: ordering and diagnosing require clinician approval, no exceptions.
That combination, Advocate Health’s strict human oversight and the token-budgeting discipline ECU Health had to build from scratch, is why efficiency and governance can’t be separate conversations in healthcare AI. An organization that solves only one is still exposed on the other, safe on cost but limited on oversight, or vice versa.
What efficient, governed AI looks like in healthcare
Pager Health℠ AI is built on a simple principle: efficiency as a design choice, not an afterthought, gets members to the right care, in the right place, at the right time, fast.
That principle pays off twice:
For health plans: members are routed correctly the first time, closing care gaps before they become claims and reducing total cost of care, the outcomes a health plan actually measures
For members: nobody waits on a slow, over-processed response to a routine question, and the AI isn’t burning excess tokens to get there
Speed for the member and cost discipline for the plan come from the same design decision, not two separate ones.
The same principle applies again the moment a conversation turns clinical. Rather than reasoning further or generating more to work through a complex case itself, AI recognizes the moment and hands off to a licensed clinician immediately, with full context attached. That hand-off is a second efficiency gain layered on the first: no additional processing gets spent trying to solve something only a clinician should decide, so token spend stops climbing at exactly the moment more computation would’ve added cost without adding safety.
This design extends across Pager Health’s care navigation and wellness platforms, which are built on advanced AI, grounded in clinical governance, and designed specifically for healthcare organizations.
An efficient model keeps members engaged long enough to finish, and knows the right moment to bring in a person instead of stretching the conversation past its usefulness. That shows up in three concrete ways:
Completed assessments: members finish the conversation instead of abandoning it
Closed care gaps: engagement that moves the actual measures rather than just logging a dial-out
Avoided escalations: the right cases reach a clinician, and everything else resolves cleanly on its own
Healthcare organizations invest in outcomes. Pager Health manages the model complexity.
That same discipline, conversational and clinical, is engineered in, not tuned later, at the model-call level:
The right model for every job. Routine, turn-by-turn work runs on a fast, lightweight model, while deeper clinical reasoning gets a stronger one. The one deliberate exception is emergency detection, which always runs on the strongest model available, because safety outranks cost, without exception.
Every call has a budget. Token limits are sized to the task rather than an unbounded default.
No wasted round trips. Responses are schema-enforced, so malformed output fails fast instead of burning a retry, and repeated prompts are cached rather than reprocessed from scratch.
Lean context, lazy work. Agentic AI works from lean, structured context instead of full transcripts, retrieval pulls back only the handful of results that matter, and detailed clinical hand-off summaries generate only the moment someone actually needs to see them.
The takeaway for healthcare tokenomics
The AI industry is still figuring out how to price and build structure around efficiency, and healthcare is experiencing that shift firsthand. The organizations getting it right are asking two questions on every interaction: how much model does this actually need, and who steps in when it’s not enough.
