Franciscan Health’s CTO can forecast his cloud bill to the month, but he can’t price a new AI agent until the invoice arrives. Here is one reason, drawn from Medicare’s own claims. An agent that runs every patient-day is billed on the average hospital stay, and the average sits well above the typical stay. There are 520 diagnosis groups below, one dot each.
how far the bill lands above a plan built on the typical stay
…
Plan on the typical stay. Get billed on the real one.
Each dot is one Medicare DRG, sized by 2023 discharges. Left to right is length of stay on a log scale: every patient-day is another day the agent reads the chart and writes. Press the red button. Each dot slides from CMS’s geometric mean stay (the typical patient) to its arithmetic mean (what the days actually add up to). Every dot moves right, because long stays pull the average up. Hover any dot.
medical DRGsurgical DRGdot area = Medicare discharges · dashed line = discharge-weighted mean stay
Discharges–
Patient-days–
Days per stay–
vs. plan–
Patient-days = Σ discharges × mean length of stay for that DRG. “Typical” = CMS geometric mean LOS. “Average” = arithmetic mean LOS. Medicare fee-for-service, IPPS hospitals only.
What moves, and what doesn’t
The newsletter’s example was a heart-failure note running far longer than a cellulitis note. CMS can see part of that: …. It can’t see the rest. Every public dataset stops at the length of the stay. How long each day’s note runs, how many chart pulls it takes, and how often the agent retries are not in any file CMS publishes. Those are the variables a clinician knows and a FinOps spreadsheet doesn’t.
Where the days are
…
Where this dataset is thin — read before quoting
Stay length is a proxy. It assumes the agent runs once per patient-day with a fixed cost. A real agent’s cost per day also depends on the length of the chart, which grows as the stay goes on. The explorer lets you model that.
Medicare fee-for-service only. This covers traditional Medicare inpatient stays at IPPS hospitals. Medicare Advantage, commercial, Medicaid, critical-access and psychiatric hospitals are not included.
Suppression. CMS drops hospital-DRG cells with 10 or fewer discharges, so rare DRGs are undercounted.
Two vintages, joined. The discharges are from calendar 2023. The length-of-stay figures come from FY2025 Table 5, which CMS builds from FY2023 claims. DRGs that were renumbered or retired in FY2025 (for example, drug-eluting stent DRGs 246–247 and several spinal fusion DRGs) drop out of the join. That removes 131,726 of 4,960,325 discharges (2.7%).
The 31% is a ratio, not a forecast. It shows how far the average stay exceeds the typical one in this case mix. Your hospital’s gap depends on your own case mix and outliers.