The numbers tell both halves of the story. 70% of healthcare organizations are now actively using AI, up from 63% the year before, and 85% of management-level respondents say AI has helped increase their annual revenue. At the same time, only 24% are achieving ROI across multiple use cases, a decline from the previous year. Both can be true, and the gap between them is where most healthcare leaders are stuck.
The models aren't the problem. The budget isn't the problem. The problem is that most operators can't tell which of their AI tools are actually working, because they never set up the measurement to find out.
What does it actually cost?
Most operators answer this by looking at the sticker price. That's almost never the real cost. Underneath the license: integration, training the team, the productivity dip while everyone learns it, the ongoing time someone spends checking outputs, then the run cost that climbs as usage scales.
The common mistake is treating AI as a fixed purchase rather than an operating model. So before signing, add up every category over twelve months. If the total lands at 2-3x the sticker price, that's normal. What's not normal is signing without having done the math.
Can the number this AI will move be named?
The biggest failure mode in operational AI is accepting benefits that sound good but can't be measured. "It'll save the team time" is a hope. "It'll cut emergency reorders from 12 a month to 3" is a benefit: it has a metric, a number, and a starting point. A real benefit has four things: a metric, a current value, a target, and a way to measure it. Miss one and it's theoretical.
The comparison hierarchy
Weakest: before / after
Measure last quarter, deploy, measure next quarter. Everything else that moved in the window lands in the "after" number.
Stronger: holdout
Roll out to six locations, hold six as a control: same season, same demand, same supplier shifts.
Strongest: staggered
Half the team this month, half next. The difference between groups is the cleanest read an operator gets.
Compared to what, exactly?
"The AI locations cut emergency reorders 40% while the others didn't move" is something a before-and-after number can never give you: proof it was the tool.
The say-do gap, 2026
Sources: NVIDIA State of AI in Healthcare 2026 · KPMG Global Tech Report 2026.
When will the result be read?
Operational AI rarely shows its value in 30 days, and rarely fails visibly in 30 days either. Early reads tell you whether the tool is being used; late reads tell you whether it's changing the business. Decide the read points (30, 90, 365 days) before you deploy, and write down what each one needs to show.
The tool decides how high you can climb. The measurement decides whether you ever get there.
The bottom line: four questions with real measurement behind them
- What does it really cost over twelve months, all in?
- What specific number will it move, from what to what?
- Compared to what, ideally a group running without the tool rather than a memory of how things used to be?
- What does success look like at 30, 90, and 365 days, and did you write it down before deploying?
