“If we stand up another agent… I can’t predict,” Franciscan Health’s CTO said this week. So here is a first pass at a meter. It uses every Medicare inpatient diagnosis group (520 DRGs, 4.8M discharges) and a documentation agent whose cost you set with the workflow facts a clinician knows: runs per day, how much chart it reads, how fast the chart grows and how often it retries.
One dot per DRG. x = Medicare discharges (log). y = modeled agent cost per stay (log). Dot area = share of the total bill. Hover for detail. Cost per stay = runs/day × (1 + retries) × Σdays[(context + growth × day) × input price + output × output price].
Cumulative share of the modeled bill, with DRGs ranked from largest to smallest share. The steeper the curve, the fewer diagnoses you need to meter well.
The spread is the story
Click through the three presets. Same patients, same stays. The only differences are the workflow assumptions, and the national bill runs from … to …, a … spread. The total is not frightening. What makes it unforecastable is how widely it can vary. Christian’s FinOps analyst can’t close that range. Someone who knows how many times a day a note actually gets touched can.
Stress-test it: turn on chart growth
With growth at zero, cost scales with stay length, and sepsis is … of the bill. Now set growth to 10k tokens per day, so that day 10’s note reads ten days of chart. Cost now rises with roughly the square of the stay. Surgical infections (DRG 853) jump to second place, ECMO/tracheostomy (DRG 003) enters the top five, and the gap against a typical-stay plan widens from … to …. The single biggest unknown in this meter is how the agent handles a long chart. That is a design choice (summarize, window, retrieve), and it is visible only to someone who has read a day-12 chart.
The scary dots don’t matter
The most expensive stays per patient sit at the top left: transplants, ECMO, CAR-T. Drag the discharge filter to 100 discharges. It removes … DRGs, and the bill barely moves (…). A per-stay horror story is not a budget line. The budget line is sepsis, heart failure and pneumonia, multiplied by hundreds of thousands of stays.
Where this model is thin — read before quoting
Every token number is an assumption. No public dataset reports how many tokens a clinical agent uses per note. The defaults (3 runs a day, 30k tokens of context, 1.5k output, 10% retries, $3/$15 per million) are round placeholders. They are not any vendor’s prices or any hospital’s measurements.
Mean stay isn’t the whole distribution. Once context grows with the stay, cost depends on the square of stay length. The average of squared stays is larger than the square of the average stay, so with growth turned on even the “average stay” setting understates the bill.
Medicare fee-for-service at IPPS hospitals only. Medicare Advantage, commercial, Medicaid, observation stays, critical-access and psychiatric hospitals are not included. The one-hospital scale applies the national Medicare case mix to 10,000 discharges, and no real hospital has that exact mix.
Suppression and joins. CMS drops hospital-DRG cells with 10 or fewer discharges. DRGs retired or renumbered in FY2025 fall out of the join with the length-of-stay table, removing 2.7% of discharges.
This is a documentation agent. Prior auth, coding and appeals agents are metered per transaction, not per day, and have their own drivers, such as payer-portal timeouts.