Measuring ROI on AI in Healthcare: The Metrics Boards Should Actually Track

Healthcare boards aren’t debating whether to invest in AI anymore. They just want to see what those investments actually delivered. In a survey by Bain & Company and KLAS Research, published in September 2026, over 300 healthcare executives weighed in. Leaders who set formal ROI targets expect their investment to pay off — three to nearly four times what they put in. But here’s the catch: a separate 2026 survey from Qventus showed 80% of health tech leaders can’t even measure AI ROI, and only 4% say their AI projects have scaled with results they can prove.

The technology isn’t usually the problem — it’s about how things get done. In most organisations, AI implementation in healthcare runs backwards: a tool is bought, deployed and celebrated, and only afterwards does someone ask what success was supposed to look like. By then there is no baseline, the benefits are scattered across departments, and finance is left trying to turn "clinicians feel less burnt out" into a number.

This article spells out the formulas, dashboard metrics, and gives two worked examples — one for a US provider, one for a UK provider. The goal? Making it possible for a board to judge an AI investment the same way they’d evaluate any other big spending decision.

Why is healthcare AI ROI so tough to pin down? Three main reasons mess things up:

First, saving time doesn’t always save money. Say an AI scribe saves a clinician half an hour a day. That time only matters if it lets them see more patients, reduces overtime, makes agency spending drop, or helps prevent staff from quitting. If the extra time just means a longer lunch break, nothing really changes on the balance sheet.

Second, the benefits often show up in someone else’s budget. An IT team buys a tool, but maybe it helps the revenue cycle team reduce claim denials and frees up outpatient capacity. If nobody keeps track across departments, each sees the costs clearly but misses out on the upside.

Third, averages can hide what’s really happening. Take The Permanente Medical Group — they reported that ambient AI scribes saved doctors about 1,794 working days’ worth of paperwork over 63 weeks, across 2.5 million patient encounters. Pretty impressive. But spread out, that’s less than a minute saved per visit. And most of the benefits came from the top third of users — 89% of activations happened there. So, if all you see is the headline number, you can’t tell if the tool works widely or just for a handful of eager adopters.

The formulas every AI business case should use

None of these formulas is exotic. What matters is using all of them together, with inputs measured in your own organisation rather than borrowed from a vendor brochure.

ROI (%): (Total quantified benefit − Total cost of ownership) ÷ Total cost of ownership × 100

Payback (months): One-off costs ÷ (Monthly benefit − Monthly running costs)

Realised time value: Users × Hours saved per user per day × Working days × Loaded hourly cost × Realisation rate

Capacity value: Additional patients seen × Average income (or cost) per episode

Denial-prevention value (US): Annual claims × Cut in denial rate (percentage points) × Cost to rework one denial

Missed-appointment value (UK): Annual appointments × Baseline DNA (did not attend) rate × Relative reduction × Cost per missed appointment

Total cost of ownership: Licences + Integration + Data preparation + Clinical validation and safety assurance + Training + Monitoring and retraining

Honestly, the realization rate is what really matters in these models — yet, most business cases just ignore it. It’s the percentage of saved time that actually translates into something useful like extra appointments, less overtime, or not having to call in agency staff. If you assume a 100% realization rate, any tool looks amazing. But if you want to be realistic, you start low, around 20% to 40%, and adjust once you see how things work in the real world.

You need that same careful approach with the ownership costs. Everybody talks about the licence fees because they're easy to spot. But cost overruns usually happen with integration — especially connecting to the electronic health record, sorting out the data, and keeping an eye on things after launch. Anyone who’s ever built healthcare software — whether it’s an EHR, telehealth platform, or AI diagnostics, like the stuff Intellectsoft builds — gets this. Hooking up a model to the records it needs isn’t just flipping a switch; it’s a whole project by itself. The board report should make this clear and show it as its own line item.

The metrics that belong on the board dashboard

A single ROI figure is a lagging indicator. By the time it disappoints, the cause is months old. Boards get earlier warnings by tracking four tiers of metrics, each feeding the next.

Tier	What it tells the board	Example metrics	Cadence
Financial	Is the investment paying back?	ROI vs plan, payback progress, cost per AI-assisted transaction, net benefit to date	Quarterly
Operational capacity	Is saved time becoming capacity?	Patients per clinician per session, documentation time per encounter, clinic utilisation, turnaround times	Monthly
Quality and safety	Is the AI safe and accurate?	Correction rate on AI output, AI-linked incidents, clinician override rate, patient experience scores	Monthly
Adoption and model health	Is the tool actually being used, and is it still performing?	Active users vs licences, use per user, accuracy vs launch baseline, drift alerts	Monthly

Boards often overlook the fourth tier, even though it actually shapes the other three. Let’s face it — a tool that only 30% of licensed clinicians use can’t possibly back up a business case meant for everyone. Same goes for any model that loses accuracy when patient mix or coding rules shift; those quiet mistakes chip away at your savings long before finance catches on.

US vs UK: same AI, different money

Now, the same AI tool plays out differently in the US and the UK, all because the money flows in different ways. In the US, where insurers pay providers for every service, AI creates value mostly by boosting revenue — more billable visits, fewer denied claims, and quicker payments. But for the NHS in the UK, where budgets are set and the real crunch is cutting down waitlists, AI’s value turns up as more capacity and lower costs. UK private providers are a bit of a blend — they sit somewhere in the middle.

Metric area	US provider	UK NHS provider
Primary value driver	Revenue capture and margin	Capacity, waiting times and cost avoidance
Revenue and activity	Initial denial rate, clean claim rate, days in accounts receivable	Clinical coding accuracy, recorded elective activity
Patient access	No-show rate, new-patient lead time	DNA rate, referral-to-treatment waits
Workforce	Overtime, locum and human scribe spend, physician turnover	Bank and agency spend, vacancy rates
Documentation	EHR time per encounter, after-hours charting	Documentation time per encounter, share of time in direct care
Regulatory gate	HIPAA; FDA where software is a medical device	UK GDPR, DTAC (Digital Technology Assessment Criteria), DCB0129/0160 clinical safety standards; MHRA where software is a medical device
Useful benchmark	Nearly 15% of claims to private payers initially denied; $43.84 average cost to contest each one	8 million missed outpatient appointments a year, costing £1.2bn; 6.4% DNA rate

Worked example 1: a US multi-specialty group

Take a physician group rolling out two tools at once: an ambient AI scribe for 150 physicians and denial-prevention AI that checks 400,000 claims a year before submission. The inputs below are illustrative assumptions, not client data. Swap in your own.

Line	Calculation	Annual value
Scribe: realised time value	150 physicians × 0.5 hrs/day × 220 days × $150/hr loaded cost × 40% realisation	$990,000
Denial prevention	400,000 claims × 3-point cut in denial rate (15% → 12%) × $43.84 rework cost	$526,080
Total annual benefit		$1,516,080
Scribe licences	150 × $250/month × 12	$450,000
Denial-prevention platform	Annual subscription	$200,000
EHR integration	One-off	$150,000
Data preparation and validation	One-off	$100,000
Training and change management	One-off	$50,000
Monitoring and governance	Annual	$75,000
Total cost of ownership, year one		$1,025,000
ROI, year one	($1,516,080 − $1,025,000) ÷ $1,025,000	47.9%
Payback	$300,000 ÷ (($1,516,080 − $725,000) ÷ 12)	4.6 months

The model is deliberately conservative. It ignores revenue recovered from denials that would otherwise have been written off, and it doesn't count extra visits made possible by the freed time. Both are real upside, but they should be added only once they have been measured.

Worked example 2: a UK NHS trust

Now take an NHS trust deploying ambient voice technology (AVT) for 120 clinicians across its emergency department and outpatient clinics, alongside AI that predicts likely non-attendance and triggers targeted reminders across 300,000 outpatient appointments a year.

The time-saving figure comes from the NHS evaluation led by Great Ormond Street Hospital, which covered 192 clinicians across nine sites. It found a 51.7% cut in documentation time, about 47 minutes per clinician per shift, and a 13.4% increase in emergency department capacity per shift. The 30% reduction in missed appointments matches the result NHS England reported from a pilot at Mid and South Essex NHS Foundation Trust. The £150 cost per missed appointment is derived from NHS England's figure of £1.2bn across 8 million missed appointments.

Line	Calculation	Annual value
AVT: realised time value	120 clinicians × 0.783 hrs (47 min) per shift × 200 shifts × £60/hr blended loaded cost × 40% realisation	£451,200
Missed appointments avoided	300,000 appointments × 6.4% DNA rate × 30% reduction × £150	£864,000
Total annual benefit		£1,315,200
AVT licences	120 × £150/month × 12	£216,000
DNA prediction service	Annual subscription	£120,000
EPR (electronic patient record) integration	One-off	£120,000
Data protection impact assessment (DPIA), clinical safety (DCB0129/0160) and DTAC assurance	One-off	£60,000
Data preparation	One-off	£60,000
Training	One-off	£40,000
Monitoring and governance	Annual	£50,000
Total cost of ownership, year one		£666,000
ROI, year one	(£1,315,200 − £666,000) ÷ £666,000	97.5%
Payback	£280,000 ÷ ((£1,315,200 − £386,000) ÷ 12)	3.6 months

Two cautions apply. First, the £60 hourly rate is an assumed blend of doctors and nurses. For context, the top of NHS Agenda for Change Band 5 is £39,043 a year in 2026/27, before employer on-costs, so a nursing-heavy rollout would carry a lower rate. Second, most of the missed-appointment value is capacity, not cash. It becomes real only when the freed slot goes to someone on the waiting list.

The downside case is the one that matters

Both examples look healthy. Now halve two assumptions: the realisation rate and the size of the denial or DNA improvement.

Scenario	US group: ROI	UK trust: ROI
Base case: 40% realisation; 3-point denial cut / 30% DNA cut	47.9%	97.5%
Downside: 20% realisation; 1.5-point denial cut / 15% DNA cut	−26.0%	−1.3%

A strong case turns into a loss in one example and break-even in the other. That isn't an argument against investing. It shows which two numbers the board should scrutinise and measure first. Every AI business case should arrive with a downside scenario and agreed stop criteria, such as pausing a rollout if active use is below 60% of licences after 90 days.

Five ways AI business cases overstate returns

1. Counting saved time at 100% realisation. Time only becomes money when someone redeploys it.

2. Using vendor case studies as the baseline. Your documentation times, denial rates and DNA rates are the only baseline that counts.

3. Leaving integration, assurance and monitoring out of the cost. These lines often rival the licence fee in year one.

4. Extrapolating from enthusiastic pilot sites. Early adopters use tools more heavily than the median clinician will.

5. Ignoring model drift. Accuracy measured at launch decays as patient mix, coding rules and workflows change. Without monitoring, benefits shrink unnoticed.

Building a board pack that holds up

The organisations that prove AI returns aren't the ones with the biggest budgets. They treat measurement as part of the build. In practice that means a few habits:

• Capture eight to twelve weeks of baseline data before go-live.

• Agree the realisation rate with finance before launch, not after.

• Report adoption, quality and capacity monthly, and financial ROI quarterly.

• Set scale-up and stop thresholds in advance.

• Reforecast ROI at six and twelve months using measured data, and retire assumptions as evidence replaces them.

Run the formulas above with your own numbers before the next AI proposal reaches the board. If the case still stands at a 20% realisation rate, it is probably worth backing. If it only works at 100%, it is a pilot, not an investment.

Frequently asked questions

What is a good ROI for AI in healthcare?

There is no universal benchmark, because returns depend on the use case, the payment model and how much of the saved time is redeployed. In the Bain & Company and KLAS Research survey, executives with formal thresholds expected three to almost four times their original investment. A more useful test for a board is whether the case still clears its own hurdle rate in the downside scenario, not just the base case.

How quickly should a healthcare AI project pay back?

Expectations are tightening. In the Qventus 2026 survey, 74% of health system technology leaders said they need to demonstrate ROI within one year. Operational tools such as ambient scribes, denial prevention and appointment reminders can pay back within months under realistic assumptions. Clinical AI that needs validation and regulatory work usually takes longer.

How do you measure the ROI of an AI scribe?

Measure documentation time per encounter for several weeks before go-live, then again afterwards for the same clinicians. Multiply the difference by encounter volume and loaded hourly cost, then apply a realisation rate for how much of that time turns into extra appointments or reduced overtime. Track use per clinician alongside it, because a few heavy users can flatter the average.

What is a realisation rate, and why does it matter so much?

It is the share of time saved by AI that becomes measurable value, such as extra patients seen, fewer paid overtime hours or fewer agency shifts. It matters because time saved is not cash saved. As the downside table shows, halving it can turn a healthy ROI negative.

Should NHS organisations measure AI ROI differently from US providers?

The formulas are the same, but the value drivers differ. US providers mostly see returns as revenue: more billable visits, fewer denied claims and faster collection. NHS organisations mostly see capacity and avoided cost, such as shorter waiting lists, fewer missed appointments and lower agency spend. NHS boards should therefore check that freed capacity is actually used, not just reported.

Which AI costs are most often left out of the business case?

The usual omissions are integration with the EHR or EPR, data preparation, clinical safety and information governance work, staff training, and ongoing monitoring and retraining. In the UK that also includes a DPIA, DCB0129/0160 clinical safety documentation and DTAC assurance. In year one these lines can rival the licence fee.

How often should the board review AI performance?

Adoption, quality and capacity metrics are worth reviewing monthly, because they give early warning. Financial ROI is better reviewed quarterly, with a full reforecast at six and twelve months based on measured data rather than the original assumptions.