# Estimating and validating ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → with realistic assumptions
A vendor tells your CFO that ambient documentation will save each physician two hours a day. Multiply that across 300 physicians and the slide deck shows tens of millions in annual value. The CFO signs. Eighteen months later, finance cannot find the savings in any budget line. This is the single most common failure in healthcare AI: confusing a vendor's theoretical ceiling with what a real hospital captures after friction.
This lesson builds an ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → model for ambient documentation (AI that listens to a clinical visit and drafts the note automatically, sometimes called an "AI scribe") and shows exactly where promised value leaks out.
A clinician wears a phone or badge mic. The AI transcribes the visit, then a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.Voir la définition complète → drafts a structured clinical note into the EHR (Electronic Health Record, the digital patient chart such as Epic or Oracle Health). The clinician edits and signs.
Real vendors here include Abridge, Nuance DAX Copilot (Microsoft), Suki, and Ambience. This is one of the most widely deployed generative AI use cases in US hospitals as of 2026, precisely because the workflow is contained and the pain (documentation burden and burnout) is severe.
The value thesis is simple: give clinicians time back. The question is how much time, and what that time is worth.
Any honest ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → model distinguishes three very different figures.
1. Vendor-promised savings: the marketing ceiling, measured in ideal pilots with motivated early adopters.
2. Gross captured time: the actual minutes saved per encounter in your setting.
3. Monetizable value: the fraction of that time you can convert into dollars or measurable outcomes.
Most failed business cases collapse #1, #2, and #3 into one number. They are rarely within 3x of each other.
Published studies and health-system pilots commonly report documentation time reductions in the range of roughly 20 to 40 percent, or a few minutes per encounter, not the "two hours a day" headline. Treat any specific figure as an estimate that varies by specialty and site. A primary care visit with heavy history differs from a 7-minute follow-up.
For a peer-reviewed anchor rather than vendor claims, the AMA's physician burnout and documentation research is a useful free starting point.
Let's model conservatively for a 400-bed hospital's affiliated outpatient clinics.
Assumptions (all flagged as illustrative estimates):
Gross time saved per year:
200 clinicians
x 16 encounters/day
x 220 days
x 3 minutes
= 2,112,000 minutes
= 35,200 clinician-hours/yearThat looks enormous. Now we apply the friction that turns hours into dollars.
Time saved is not money until something changes. Here is where the vendor deck goes quiet.
Not every licensed clinician uses the tool, and not from day one. Realistic patterns:
If only 65 percent of your 200 seats are truly active, your effective base is 130 clinicians, not 200. That alone cuts the raw number by 35 percent.
The AI drafts; the clinician must still review and correct. Early on, editing can eat much of the promised savings, especially for complex specialties or hallucinated details that must be caught for patient safety. Net savings, not gross transcription time, is what counts.
This is the decisive step. Saving a physician 20 minutes a day does not automatically produce revenue or reduce cost. It only does so if one of these happens:
Time returned as "less burnout" is real and valuable but shows up as retention and quality, not a line item finance can invoice.
Let's monetize honestly. We take the 130 active clinicians and assume each converts saved time into 2 additional encounters per day where demand exists (a strong but plausible assumption for a busy system), for half of clinicians (the rest bank the time as reduced burnout).
Throughput value (illustrative estimate):
65 clinicians (half of 130)
x 2 extra encounters/day
x 220 days
x $60 estimated contribution margin per encounter
= $1,716,000/yearContribution margin per encounter varies enormously; $60 is a placeholder you must replace with your own payer mix.
Retention value (illustrative estimate):
Assume the tool reduces annual physician turnover by even 3 avoided departures, at a commonly cited replacement cost estimate of several hundred thousand dollars each (recruitment, lost billings, onboarding). At $250,000 avoided per departure:
3 x $250,000 = $750,000/yearTotal realistic annual value: ~$2.47M
Now the cost side.
Costs (illustrative):
200 seats x $300 x 12 = $720,000/year(You pay for licensed seats, not just active ones.)
Year 1 net:
$2,470,000 value
- $720,000 licenses
- $350,000 one-time
= $1,400,000 netPositive, but roughly 6x smaller than a naive "two hours x 200 doctors" headline would suggest. That gap is the entire lesson.
🎬 [VIDEO: "The ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → of Ambient AI Scribes in Healthcare" - youtube.com - a practical walkthrough of how health systems measure documentation AI value and where estimates break down]
Vérification des acquis
1. Why does a vendor's promised time savings figure typically overstate the value a real hospital captures?
2. A hospital measures that ambient documentation saves clinicians real minutes per encounter, but finance still cannot find corresponding savings in the budget. What distinction best explains this gap?
3. What is the core reasoning error behind the failed CFO business case described in the lesson?
4. Select ALL correct answers. Which factors help explain why ambient documentation is one of the most widely deployed generative AI use cases in hospitals?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. When building an honest ROI model for ambient documentation, which practices align with the lesson's approach?
Sélectionnez toutes les réponses correctes.
A model is only as good as its inputs. Validate each driver against reality, not the slide.
Decide before rollout exactly what you will measure and how. Otherwise you will rationalize whatever you get. Core metrics:
Compare adopters to a matched non-adopter group, or use a staggered rollout (some clinics go first). This isolates the AI's effect from seasonal volume swings. Without a comparison group, you cannot claim the AI caused the change.
Clinicians who like the tool overreport savings. A 20-minute self-reported saving may be 6 minutes on the clock. Use direct measurement for at least a sample.
Usage decays, editing improves, and license prices renegotiate. An ROIROI model is a living document, not a one-time approval artifact. Rebuild it every quarter with observed numbers replacing estimates.
The classic failure: finance owns the spreadsheet but not the workflow, and clinical leaders own the workflow but never see the assumptions. The monetization step (does saved time become throughput, retention, or nothing?) requires both. If your open-slot demand is low, throughput value is near zero no matter how good the AI is, and your entire case must rest on retention and burnout.
Name the assumption out loud, then decide whether it is investable.