clinicians.build · interactive · september 1, 2026
Year Two
Mayo Clinic has 128 clinical AI solutions in practice and is approaching 500 in the pipeline. Roughly 70% are built in-house. Every dot below is one of them. The count is published. The number that actually decides whether this works is not.
Primary source: Micky Tripathi, PhD, Chief AI Implementation Officer, Mayo Clinic, to Becker's Hospital Review, Aug 31 2026
Context: Deloitte 2026 Healthcare CFO Survey · UPMC AI governance research, Aug 6 2026
We spent three years arguing about whether clinical AI works. Mayo's answer is 128 models deep and roughly 500 more coming. The argument that decides the next three years is a different one, and Tripathi named it: “The cost of maintaining these systems is far higher than I think any of us in the industry really thought.”
Three cost centres — compute, the data infrastructure underneath, and the workforce to watch the models. The third is the trap. Traditional software sits still after deployment. A model can't: the data moves, the population moves, the coding rules move, and someone has to be looking.
Every model you ship subtracts from the team that ships the next one.
628 models, filling in
Each dot is one clinical AI solution at Mayo. Solid is live in practice; hollow is in the pipeline. The darker ring marks the roughly 70% built in-house — the ones where, in Tripathi's words, “we are ultimately, for those solutions, the legal manufacturer.” Press Ship and watch the field fill.
live in practice (128)in the pipeline (~500)built in-house — Mayo is the legal manufacturer
the two numbers Mayo has never published
20
8
live in practice
128
of 628 total
engineers on watch
16.0
maintenance, not building
left to build
4.0
engineers free
team locked
80%
share on maintenance
128 live at 8 models per engineer consumes 16 of 20 engineers. Four are left to build the next 500. Now press Ship.
80/20 lens
Two of the three numbers in that panel are yours, not Mayo's. The 128 and the ~500 are published. The engineering headcount and the models-per-engineer ratio are not — not by Mayo, not by anyone. Notice that moving both sliders does nothing to the dot field and everything to the bar. The count is the part that gets announced; the ratio is the part that decides whether year two is survivable. If you are running one model, you have one model and one engineer who can never take a vacation. The ratio is what scales, not the count.
Same gap, three chairs
The maintenance bill isn't a Mayo eccentricity. It shows up as three separately-reported numbers that are really one problem: nobody can prove the return, so nobody funds the watching.
Mayo · built in-house
~70%
Not licensed. Built. Which makes the health system the legal manufacturer, with the regulatory and monitoring duty that carries. Becker's →
CFOs · mature financial attribution
18%
Of health systems scaling generative AI, fewer than one in five can attribute a financial result to it. The CFO can't prove the return because the counterfactual was never built. Deloitte 2026 →
Systems · deployed vs. able to test
93% / 44%
93% have deployed third-party AI. 44% have anywhere to test it first. The red band is the 49-point gap where a model goes live without a sandbox behind it. UPMC research →
Mayo funds part of the monitoring with philanthropy, because — Tripathi again — “there is no line-of-sight ROI because you're just improving quality in a system that doesn't compensate for higher quality right now.” That is the whole problem in one sentence: quality is not a billable event, and the watch is a quality activity.
What this graphic can't tell you
628 is a count, not a workload. A sepsis model retrained quarterly and an ambient-scribe deployment are one dot each. Effort per model probably varies by an order of magnitude, and nothing public lets you weight the dots.
“Approaching 500 in the pipeline” is not 500 shipping. Pipeline attrition in clinical AI is high and unmeasured. Reading the hollow dots as future live dots is the most optimistic possible assumption.
The three panels below are three different surveys with different populations, questions and years. They rhyme; they are not the same denominator.
The maintenance quotes could not be independently re-verified. Becker's serves an automated bot challenge, so the four quoted passages here come from the article as filed. The 128 / ~500 figures and Tripathi's title were corroborated separately.