Investigative Research Report Radiology Leadership Series

The Algorithm Works. The Math Does Not.

An Investigation Into Radiology AI’s Return on Investment Problem

Kelly Emrick, DHSc, PhD, MBA, BSRT(ARRT)R November 2026

  • 14 figures
  • 6 tables
  • 34 references
  • The PAYS Test
Report cover: The Algorithm Works. The Math Does Not.

At a glance

Radiology AI creates real value, but most of that value lands where the buyer cannot book it. That is why only about one in four organizations can show a quantified financial return.2

24%

can show a quantified positive ROI2

14%

are reimbursed because AI ran2

2

AI services hold a payable Category I CPT code10

12%

cut radiologist or locum spend and kept the savings2

2.5%

Medicare cut to work RVUs for assumed efficiency13

The central finding

Value created is different from value captured. An AI tool earns its keep only when the organization paying for it also holds the ledger where the benefit lands.

Roughly a thousand cleared algorithms. Two payable codes.

Each dot is one FDA-cleared radiology AI device (about 1,000 shown, an approximation by the author). Gold: fractional flow reserve CT and coronary plaque analysis.10, 30

Five conclusions

  1. The headline statistic is thinner than it looks. The “three in four cannot prove ROI” figure traces to a six-day flash poll of 187 organizations, revived this fall by an AI vendor executive.1, 2
  2. The best-known positive figure is no stronger. The 451% five-year ROI comes from a model built for a platform its maker shut down in September 2025.6, 7, 8
  3. Almost nobody is paid for using AI. Radiology holds roughly 80% of FDA-cleared AI algorithms, yet only two AI services carry a payable code.2, 10
  4. Time saved rarely becomes money earned. Turnaround gains vary by tool, site, and shift.15, 16, 17, 18
  5. Medicare is already pricing in the efficiency. The payer is collecting the dividend before the buyer can.13

Where the value lands

Radiology carries the cost. The benefit flows to other departments, the payer, radiologists (as relief), and patients (as better care). Only one path returns to the ledger that paid for the tool.

Radiology pays

License, integration, IT support, training

AI tool creates value

  • Other departmentsFaster decisions in the ED, neurology, oncology
  • The payer2.5% efficiency adjustment to work RVUs
  • RadiologistsTime absorbed by volume; lower strain
  • PatientsEarlier detection, fewer missed findings
  • Radiology’s own ledgerOnly with a payer, convertible capacity, or captured follow-up

The PAYS Test in one look

PPayer

Who writes a check because this tool ran?

AAttribution

Can we prove the change came from the tool?

YYield

Does the saving convert into revenue or avoided cost that we keep?

SSunset

When do we review, and what result ends the contract?

Abstract

Radiology holds most of the cleared artificial intelligence tools in medicine, yet most organizations that use them cannot show they pay for them. This investigation traced the leading return-on-investment claims to their sources and tested them against peer-reviewed productivity studies, Medicare payment policy, and vendor financial results. It finds that the best-known pessimistic statistic and the best-known optimistic statistic both originate with parties that have a commercial interest. It also finds a structural cause that is independent of either number: radiology AI has almost no revenue line of its own, saved time rarely converts into revenue or avoided cost, and Medicare has begun reducing physician payment on the assumption that technology raises productivity. The return is real in a small number of lanes where a payer exists or freed capacity is convertible. The report proposes the PAYS Test, a four-question framework for separating AI investments from AI expenses before signing a contract.

Keywords: radiology, artificial intelligence, return on investment, reimbursement, productivity, imaging leadership, healthcare finance, technology governance

The question and the evidence trail

1. The question

Radiology spent a decade asking whether AI works. For many tools, that question is settled well enough for clinical use. A Doximity survey published November 12, 2026, found that 64% of radiologists use AI daily.5

The harder question is who can prove it pays. On October 30, 2026, The Imaging Wire reported that three out of four organizations using imaging AI cannot demonstrate a quantified positive return.1 A claim that convenient deserves scrutiny. This report followed it to its source, then widened the search.

2. How the investigation was conducted

Five bodies of evidence were examined:

  • The industry surveys behind the ROI statistics
  • Peer-reviewed and conference studies of real-world productivity
  • Medicare payment policy for AI
  • Vendor financial results and platform exits
  • Use cases with documented payback

We graded every source by its independence and method. Where a figure comes from a vendor, a press release, or a model, the text says so.

Table 1. Evidence grading used in this report
GradeType of sourceExamples in this report
APeer-reviewed study of observed resultsJACR triage study17; JMIR two-hospital study18
BRegulatory action or audited company filingCMS fee schedule rules13, 14; RadNet quarterly results25
CConference presentation or trade reporting of a studyYale and University of Washington studies15, 16
DIndustry survey released without full methodsBlack Book poll2; Croviz survey4
EVendor-reported case result or vendor-built modelSwiftMR site results1; Calantic ROI calculator6

Evidence register

Every graded source in the report. Filter by grade to see how much of the debate rests on each kind of evidence.

Showing 14 of 14 sources

Graded sources
SourceWhat it showsRef.Grade
JACR pulmonary embolism triage studyTurnaround fell in work hours only17, 33A
JMIR two-hospital lung nodule studySlower overall; one site faster by year two18A
JAMA Network Open report drafting study15.5% average efficiency gain19A
Banner MD Anderson follow-up study67 cancers among 445 navigated patients28A
CMS fee schedule rules2.5% efficiency adjustment; 2027 conversion factor cuts13, 14B
RadNet quarterly resultsSegment revenue $32.9 million; adjusted EBITDA $3.0 million25B
Mammography AI self-pay add-on$40 per exam; adoption approaching 45%26, 27B
CPT codes and CT triage add-onTwo Category I codes; up to $137.53 per eligible case10, 11B
Yale study, ARRS 2026No change in turnaround, treatment time, or length of stay15C
University of Washington, RSNA 2026Large gains for some findings, none for PE or hemorrhage16C
Black Book flash poll24% report a quantified positive ROI2D
Croviz survey30% have quantified a positive ROI4D
SwiftMR site results42% volume gain; $205,567 in two months1E
Calantic ROI calculator451% five-year ROI (791% with radiologist time)6E

How the ROI question reached this point

Key events, 2020 to 2026. Gold: platform events. Green: payment policy. Blue: evidence.

  1. 2020First AI tool earns a Medicare add-on payment
  2. Jun 2022Bayer launches the Calantic AI platform
  3. Feb 2023Bayer completes its Blackford acquisition
  4. Mar 2024JACR calculator projects 451% ROI for Calantic
  5. Apr 2025Health Tech Investment Act introduced in the Senate
  6. Sep 2025Bayer discontinues Calantic and Blackford services
  7. Nov 2025Black Book poll: 24% can show quantified ROI
  8. Jan 2026CMS 2.5% efficiency adjustment takes effect
  9. Apr 2026Croviz survey: 30% have quantified ROI
  10. Oct 2026Sopris buys Blackford; a vendor article revives the 24% figure
  11. Nov 2026CMS finalizes 2027 conversion factor cuts
Figure 1. Key events in the radiology AI ROI debate, 2020 to 2026. Author’s compilation from references 2, 4, 6, 8, 9, 12, 13, 14.

Two threads run through the timeline. One is the rise and exit of a major AI platform. The other is a payment system that has moved slowly to pay for AI and quickly to discount for it.

3. Adoption without proof

Both the pessimistic and optimistic numbers in this debate come from parties with something to sell. That is the first finding, and it shapes everything after it.

Where the “three in four” figure comes from

The Imaging Wire article was written by Jina Park, chief strategy officer of AIRS Medical, which sells the MRI enhancement product SwiftMR. One indexed version of the page carried a Partner Content tag. The article’s success stories, a 42% volume gain and $205,567 in added revenue at one imaging center, are SwiftMR customer results.1

Its central statistic comes from a Black Book Research flash poll fielded November 25 to 30, 2025. The poll had 212 respondents, and the ROI questions covered the 187 organizations with AI live or in pilots. It was released as a press release with a stated margin of error of 6 points and no published questionnaire.2

That does not make the finding wrong. A second survey, by AI governance startup Croviz, reached a similar result: 30% of 445 respondents had quantified a positive ROI. Croviz sells AI oversight, and its sample included vendor executives.4 Two imperfect surveys pointing the same way are suggestive. They are not proof.

One in four can prove the return

Share of organizations using imaging AI that report a quantified positive ROI

24%can show it
  • 24% report a quantified positive financial ROI
  • 76% do not: unquantified, neutral, negative, or never measured
  • Base: 187 organizations with imaging AI live or in pilots. Margin of error: 6 points.
Figure 2. Share of organizations able to show a quantified positive ROI from imaging AI. Source: Black Book Research RSNA Flash Poll, November 25 to 30, 2025 (reference 2).

What the poll actually shows

The details in the Black Book poll matter more than its headline. Only 14% of organizations collect reimbursement tied directly to AI. Another 39% report only indirect upside, such as higher throughput or fewer denials. Nearly half either see no attributable revenue or cannot attribute it.2

Only 14% are paid because AI ran

How organizations with imaging AI describe its revenue effect

Direct reimbursement tiedto AI14%Indirect upside only(throughput, new services,fewer denials)39%No revenue impactattributable to AI31%Cannot attribute revenue toAI16%
Figure 3. Revenue effect of imaging AI as reported by organizations using it. Source: Black Book Research RSNA Flash Poll, November 25 to 30, 2025 (reference 2).

The largest group never checks

When organizations review the financial performance of imaging AI

Formal financial reviewwithin 12 months of go-live30%Review, but later than 12months32%No defined financial reviewprocess38%
Figure 4. Timing of financial review after imaging AI deployment. Source: Black Book Research RSNA Flash Poll, November 25 to 30, 2025 (reference 2).

The poll also explains its own headline. Most organizations cannot prove a return because most never set up a way to measure one.2

Contracts renew on goodwill, not on throughput

Main reason cited by sites that renewed an AI vendor contract

Better radiologist qualityof life54%Reduced turnaround time18%More scans per reader10%Reduced downstream patientcosts10%Better diagnostic accuracy8%
Figure 5. Main reason cited for renewing an AI vendor contract. Source: Croviz survey of 445 radiology AI users, as reported by The Imaging Wire (reference 4).

Why buyers renew anyway

In the Croviz survey, 95% of AI sites renewed at least one vendor contract in the past year. The leading reason, cited by 54%, was better radiologist quality of life. Only 10% cited more scans per reader, and only 6% tracked hard metrics such as retention.4

Adoption is being driven by the people who read the studies, and the finance office is following. That pattern becomes fragile at budget time. Asked what they would cut first under financial pressure, 34% of polled leaders named new AI tools, ahead of analytics, cloud imaging, modality upgrades, and staffing.2

The optimistic number has the same problem

The most cited positive figure is a 451% ROI over five years, rising to 791% when radiologist time is counted, from a 2024 paper in the Journal of the American College of Radiology.6 It is a calculator. Its inputs came from expert interviews and literature, and its output is a projection for a modeled hospital.

The calculator was created to assess Calantic, Bayer’s AI platform.7 In September 2025, Bayer discontinued Calantic and the services of its Blackford subsidiary, ending a five-year effort.8 The platform modeled to return 451% to its customers did not survive its owner’s own business review.

What was not found

This investigation did not find an independent, audited study of observed financial return from radiology AI at a U.S. provider. The debate is being conducted with a press-release poll on one side and a vendor model on the other.

4. Follow the money

Radiology AI has almost no revenue line of its own, so nearly every return must be earned indirectly. That single fact explains most of the ROI problem.

Hundreds of cleared tools, two payable codes

Each dot is one FDA-cleared radiology AI device (about 1,000 shown, an approximation). Gold: AI services with a Category I CPT code.

  • Fractional flow reserve CT
  • Coronary plaque analysis (from January 2026)
Figure 6. FDA-cleared radiology AI devices compared with AI services holding a Category I CPT code. Source: FDA device list count and radiology share as reported in references 10 and 30; approximation by the author.

Hundreds of cleared tools, two payable codes

Radiology accounts for roughly 80% of FDA-cleared AI algorithms in medicine, on an FDA list that passed 1,300 devices in December 2025.10, 30 As of January 2026, only two AI services held a payable Category I CPT code: fractional flow reserve CT and coronary plaque analysis.10

Many vendors advertise “a CPT code.” Most of those are Category III codes, which track use and do not guarantee payment. Eric Rubin, MD, the American College of Radiology’s CPT advisor, told Radiology Business that tools performing tasks radiologists already do, such as fracture or nodule detection, are unlikely to earn their own paid codes.10

Medicare has already booked the savings

While providers wait to be paid for AI, Medicare has started paying less because of technology. The 2026 physician fee schedule applied a 2.5% efficiency adjustment to work RVUs for most non-time-based services, radiology included. CMS will repeat it every three years.13

The agency’s premise is that workflow improvements and technology adoption make physicians more productive. The provider pays for the tool, and the payer collects the efficiency in advance. The 2027 rule, finalized November 4, adds conversion factor cuts of 1.83% and 2.31%.14

Table 2. Medicare payment pathways for radiology AI
PathwayStatusLimits
Category I CPT codeTwo AI services as of January 202610Requires wide use and published evidence of clinical value
Category III CPT codeHeld by a small number of tools10Tracks use; does not guarantee payment
Inpatient add-on payment (NTAP)Up to $137.53 per eligible case for one CT triage product in fiscal 202711One product, inpatient only, three-year window, documented use
Health Tech Investment ActIntroduced in Senate (April 2025) and House (November 2025)12, 32Not enacted; trackers list both bills as introduced
Patient self-payUsed for mammography AI at $40 per exam26Depends on patient willingness to pay

Radiology carries the cost. The benefit flows to other departments, the payer, radiologists (as relief), and patients (as better care). Only one path returns to the ledger that paid for the tool, and it opens only under specific conditions. The value flow diagram (Figure 7) appears on the At a Glance tab.

5. The productivity paradox

AI often saves minutes, but minutes become dollars only when they turn into billable exams or avoided costs. The published evidence shows how often that conversion fails.

Faster reports that changed nothing downstream

Yale researchers studied 484 emergency patients with serious pulmonary embolism before and after AI triage. Report turnaround was just over one hour in both periods. Length of stay was about 5.9 days before AI and 6.1 days after.15

Presenter Michelle Li concluded that report speed was probably not the limiting factor in patient outcomes.15 A tool cannot return value by speeding up a step that was never the bottleneck.

Gains that depend on the tool, the site, and the shift

Eleven tools, one health system, very different results

Reduction in median report turnaround after AI triage, by finding (five of eleven tools reported)

Endotracheal tubemalposition76%Rib fracture53%Brain vessel occlusion29%Pulmonary embolism0%Intracranial hemorrhage0%
Figure 8. Change in report turnaround by finding after system-wide AI triage across three hospitals and more than 742,000 exams. Source: University of Washington, RSNA 2026 preview via AuntMinnie (reference 16). Zero indicates no improvement reported.

The same tool, two different results

Mean report turnaround in minutes for PE-positive CT pulmonary angiography, one academic center, 11,252 exams

Work hours68.946.7Off hours44.842.0Before AI triageAfter AI triage
Figure 9. Report turnaround before and after AI triage, by time of day. Source: JACR, January 2026 (references 17 and 33).

One hospital needed two years. The other saw no gain.

Change in report drafting time per chest CT versus controls (difference in differences), 185,044 reports

Both sites pooled+0.86 minSite B, first year (not significant)+0.90 minSite A, first year-0.70 minSite A, second year-2.66 minFaster with AISlower with AI
Figure 10. Change in report drafting time with AI at two hospitals. Source: Journal of Medical Internet Research, 2026 (reference 18). Site A: Beijing Anzhen. Site B: Tsinghua Changgung.
Table 3. Real-world productivity studies reviewed
StudySettingResultGrade
Yale, ARRS 202615PE triage, 484 emergency patientsNo change in turnaround, treatment time, or length of stayC
University of Washington, RSNA 20261611 triage tools, 3 hospitals, 742,000 examsLarge gains for some findings, none for PE or hemorrhageC
JACR, 202617, 33PE triage, 11,252 exams, one centerGain in work hours onlyA
JMIR, 202618Lung nodule AI, 185,044 reports, 2 hospitalsSlower overall; one site faster by year twoA
JAMA Network Open, 202519In-house report drafting, about 24,000 reports15.5% average efficiency gain, up to 40% for some readersA

The pattern is consistent. Results vary by algorithm, by hospital, and by time of day, and one site needed two years to see savings. A buyer who measures at six months may be measuring the learning curve.

Why saved time does not reach the ledger

  • The saved time is too small to redeploy. A minute or two per report does not eliminate a shift or a locum contract. Only 12% of polled organizations cut radiologist or locum spending and kept the savings.2
  • The benefit lands in another department. A faster stroke read helps neurology and the emergency department. Radiology carries the license and integration cost.1
  • The time is absorbed, not sold. Radiologists use it to keep up with rising volume or to recover from overload. That has value, but nobody invoices for it.

Black Book founder Doug Brown put it plainly: buyers “can’t point to a clean line on the P&L” showing the tools pay them back.3

Saved time rarely becomes saved money

Staffing effect of imaging AI reported by organizations using it

Reduced radiologist orlocum spend and kept thesavings12%Avoided hiring orredeployed staff, no netcost reduction24%No measurable staffingimpact42%Too early to tell or nottracked22%
Figure 11. Staffing effect of imaging AI. Source: Black Book Research RSNA Flash Poll, November 25 to 30, 2025 (reference 2).

6. The vendor side of the ledger

If customers were banking large returns, the companies selling the tools would be thriving. The seller’s side of the ledger tells the same story as the buyer’s.

A decade of investment against one year of sales

Medical imaging AI, US$ billions. The two bars measure different things and are shown for scale only.

Venture capital invested(cumulative since 2015)about $3.6BGlobal market revenue(single year, 2024)about $0.75B
Figure 12. Cumulative venture investment in medical imaging AI compared with one year of market revenue. Source: Signify Research, as reported in references 20 and 21.

Revenue up 32.6%, segment earnings down 11.8%

RadNet Digital Health segment, US$ millions. The 2025 EBITDA value is derived from the reported 11.8% decline.

Third quarter 2025$24.8Mabout $3.4MThird quarter 2026$32.9M$3.0MSegment revenueSegment adjusted EBITDA
Figure 13. RadNet Digital Health segment revenue and adjusted EBITDA. Source: RadNet third quarter 2026 results (reference 25).

The market is small

Signify Research put global medical imaging AI revenue at about $749 million in 2024. Venture investors have put almost $3.6 billion into the sector since 2015.20, 21 Signify analysts wrote in early 2025 that only a small handful of vendors could claim break-even status.22

A platform owner walked away

Bayer exited the platform business in September 2025, and Sopris Capital bought Blackford in October 2026 with plans to fold it into Azra AI.8, 9

Providers are buying the vendors

RadNet agreed to pay up to $270 million for Gleamer, and Radiology Partners bought Cognita for $80 million.23, 24 The organizations most willing to own radiology AI also own the scanners, the radiologists, and the patient relationship.

Thin margins, foggy pricing, and a new measurement market

Even the scaled seller earns thin margins. RadNet’s Digital Health segment reported third quarter 2026 revenue of $32.9 million, up 32.6%. Segment adjusted EBITDA was $3.0 million, down 11.8%, or about 9% of segment revenue.25

Pricing adds to the fog. In the Black Book poll, 36% of buyers paid per study, 25% received AI bundled into modality or PACS contracts, and 20% held enterprise licenses.2 Bundled AI has no separate price, so it can have no separate return.

A new kind of company has appeared to fill the gap. TestDynamics raised $9 million in November 2026 for a platform that benchmarks imaging AI, and Croviz sells governance.4, 31 A market for measuring AI is growing because measurement was never built in.

7. Where the return is real

Radiology AI pays where someone is billed for it or where the buyer captures the capacity it frees. Four lanes have evidence behind them, though the quality varies.

Table 4. Use cases with a documented or plausible financial return
LaneExampleEvidenceStrength
A patient or payer pays directlyMammography AI add-on at $40 out of pocket, with national adoption approaching 45%26, 27Company reporting and earnings calls BStrong on revenue
A code or add-on existsFFR-CT, coronary plaque analysis, CT triage add-on10, 11CMS and CPT actions BStrong, narrow
Saved scanner time becomes billable examsMRI acceleration: 42% volume gain and $205,567 in two months at one center1Vendor reported, unaudited EPlausible, unverified
Findings convert into captured follow-up careBanner MD Anderson: 961,300 reports screened, 445 patients navigated, 67 cancers diagnosed28Peer-reviewed clinical study, no dollar figures AStrong clinically, unpriced

What separates these lanes

First, each has a constraint that money was already waiting behind. A full MRI schedule converts minutes into exams. A screening patient will pay $40 for a second look. An incidental finding becomes an oncology referral.

Second, each pairs the algorithm with an operational change. The Banner program worked because staff navigated the flagged patients into multidisciplinary care. The software found the findings, and the workflow captured them.28

A 2026 review in HealthManagement.org reaches the same conclusion from the literature: profitability comes mainly from downstream clinical pathways, and time savings count only when converted into billable activity or avoided costs.29

Payer lane estimator

Simple arithmetic on two payment figures cited in the report. An illustration, not a forecast.

Add-on payment ceiling at $137.53 per case: $137,530.00

Self-pay revenue at $40 per exam: $360,000

The add-on figure is a per-case maximum for one CT triage product in fiscal 2027, with documented use.11 The $40 fee and an adoption rate approaching 45% are RadNet’s reported figures; your results will differ.26, 27 Neither line subtracts the cost of the tool.

8. The case against this report’s conclusion

The strongest objection to this report is that financial ROI is the wrong test for a clinical tool. That argument deserves a fair hearing, and parts of it are correct.

“PACS never proved its ROI either.”

Imaging Wire writer Brian Casey made this comparison when covering the Croviz survey.4 Radiologists drove PACS adoption because it made the work better, and nobody would go back to film. AI may follow the same path.

“Unmeasured is not the same as absent.”

The polls count organizations that cannot show a return. They do not show that returns are negative. In the Black Book poll, 59% reported faster turnaround and 44% reported better detection.2

“The alternative to AI is not the status quo.”

With volumes rising and radiologists scarce, the realistic comparison may be AI versus unread studies, burned-out readers, or costly locums. Retention is hard to price and expensive to lose.

“Quality has its own return.”

Catching cancer early or avoiding a missed finding carries clinical and legal value. In November 2026, a jury awarded $7.8 million over a lung mass missed on a breast MRI.31

The reply

These points are sound, and they change the conclusion less than their advocates suggest. Each describes a benefit that can be measured: turnaround, retention, locum spend, claims. A benefit that leaders decline to measure cannot be defended at budget time.

9. The PAYS Test

The evidence points to one governing idea: value created is different from value captured. The PAYS Test turns that idea into four questions to answer before purchase and again at renewal.

It is a framework proposed in this report, drawn from the evidence, and it has not been validated in practice.

Score a tool

Answer each question for one AI tool. The result follows the scoring rule in Table 6.

P
Payer

Who writes a check because this tool ran?

A passing answer: A named source: a code, an add-on payment, a patient fee, or a contract rate

14% collect AI-tied reimbursement2

A
Attribution

Can we prove the change came from the tool?

A passing answer: A baseline of 6 to 12 months captured before go-live, with a comparison group where possible

38% have no financial review process2

Y
Yield

Does the saving convert into revenue or avoided cost that we keep?

A passing answer: A named conversion: added exam slots, a canceled locum contract, a captured referral

12% kept staffing savings2

S
Sunset

When do we review it, and what result ends the contract?

A passing answer: A review date within 12 months and written exit criteria

30% review within 12 months2

Result

Answer all four questions

0 of 4 answered.

The PAYS Test

Four questions to answer before purchase and again at renewal, with the poll evidence behind each

PPayer

Who writes a check because this tool ran?

14% collect AI-tied reimbursement

AAttribution

Can we prove the change came from the tool?

38% have no financial review process

YYield

Does the saving convert into revenue or avoided cost that we keep?

12% kept staffing savings

SSunset

When do we review it, and what result ends the contract?

30% review within 12 months

Figure 14. The PAYS Test for radiology AI investment decisions. Framework proposed by the author; percentages from the Black Book poll (reference 2).
Table 5. The PAYS Test in detail
ElementQuestionWhat a passing answer looks likeEvidence behind it
PayerWho writes a check because this tool ran?A named source: a code, an add-on payment, a patient fee, or a contract rate14% collect AI-tied reimbursement2
AttributionCan we prove the change came from the tool?A baseline of 6 to 12 months captured before go-live, with a comparison group where possible38% have no financial review process2
YieldDoes the saving convert into revenue or avoided cost that we keep?A named conversion: added exam slots, a canceled locum contract, a captured referral12% kept staffing savings2
SunsetWhen do we review it, and what result ends the contract?A review date within 12 months and written exit criteria30% review within 12 months2
Table 6. Scoring the PAYS Test
ResultClassificationLeadership action
Passes all fourInvestmentHold it to a financial return and report that return annually
Fails Payer and Yield, passes Attribution and SunsetQuality or workforce expenseBudget it as a cost of safe, sustainable reading and judge it on the clinical or retention metric it was bought to move
Fails Attribution and SunsetUnmanaged spendDo not buy. Without a baseline and a review date, no one will ever know what it did

The capture question in one line

For any claimed benefit, ask where it lands. If it lands in the emergency department, the payer’s rate file, or the radiologist’s evening, the imaging budget will not see it. That may be acceptable. It should be a decision, made in advance and stated in the business case.

10. What leaders should do now

Leaders heading to RSNA 2026 on November 29 will be shown hundreds of AI products. Six actions follow from the evidence.

Leadership action checklist

Check off each action as your organization completes it. Progress is saved in this browser.

0 of 6 actions complete

The payment question

Hospital boards and finance leaders should also press the payment question in Washington. Until Medicare pays for AI services the way it now discounts for assumed efficiency, providers will fund the technology and payers will collect the gain.

11. Limits of this investigation

This is a document-based investigation, and readers should weigh it accordingly.

No original interviews or records.

The findings rest on published studies, filings, trade reporting, and press releases. Confirming them would require interviews with health system finance leaders and access to contracts and internal ROI reviews.

The core surveys are not peer reviewed.

The full Black Book and Croviz reports were not available. Their figures come from the press release and trade coverage.

Some sources were read in summary only.

The JACR calculator paper and the HealthManagement.org review were assessed from abstracts and secondary reporting.

Two studies are conference previews.

The University of Washington and Yale results have not yet appeared as full peer-reviewed papers.

Vendor case figures are unaudited.

The SwiftMR and RadNet numbers are reported by the companies that benefit from them.

Some values are approximations.

The device count in Figure 6 and the 2025 EBITDA value in Figure 13 are derived by the author from reported figures.

Legislative status may have changed.

The status of the Health Tech Investment Act comes from bill trackers and should be confirmed before publication.

Absence of evidence has limits.

This search did not find an independent, audited study of observed ROI. One may exist.

References

All 34 sources cited in the report, with links. Filter by topic.

Showing 34 of 34 references

  1. Park J. Radiology AI Has an ROI Problem. The Imaging Wire. October 30, 2026. Open source
  2. Black Book Research. Black Book RSNA Flash Poll: 3 in 4 Imaging Leaders Have No Clear ROI for AI. Press release. December 2025. Open source
  3. Radiology Business. Most organizations using AI for radiology are unsure of its ROI. December 2025. Open source
  4. Casey B. AI's ROI Paradox. The Imaging Wire. April 12, 2026. Open source
  5. AuntMinnie. Daily AI use high among radiologists, may help hone skills. November 2026. Open source
  6. Bharadwaj P, Nicola L, Breau-Brunel M, et al. Unlocking the Value: Quantifying the Return on Investment of Hospital Artificial Intelligence. Journal of the American College of Radiology. 2024;21(10):1677. Open source
  7. Radiology Business. Radiology artificial intelligence ROI calculator demonstrates 'substantial' benefits at 5-year mark. March 2024. Open source
  8. AuntMinnie Europe. Major shake-up looms for AI platform sector. September 2025. Open source
  9. Beri. Sopris Buys Blackford From Bayer, Silent on Its 45 AI Vendors. October 2026. Open source
  10. Radiology Business. Radiology dominates FDA-cleared AI, but reimbursement lags far behind. December 2025. Open source
  11. MarketScale. Medicare sets $137.53 maximum add-on payment per eligible inpatient case for CT triage AI. August 30, 2026. Open source
  12. AuntMinnie. Senators introduce medical AI reimbursement bill. April 2025. Open source
  13. American College of Radiology. Decoding the 2026 Medicare Physician Fee Schedule Final Rule. ACR Bulletin. Open source
  14. AuntMinnie. CMS final rule draws mixed response. November 2026. Open source
  15. AuntMinnie. ARRS: AI does not improve efficiency or outcomes in PE care. April 2026. Open source
  16. Ridley EL. AI triage algorithms yield varied results in real-world use. AuntMinnie. November 11, 2026. Open source
  17. Impact of Artificial Intelligence Triage on Radiologist Report Turnaround Time: Real-World Time Savings and Insights From Model Predictions. Journal of the American College of Radiology. 2026;23(1):63. Open source
  18. Effect of AI-assisted lung nodule diagnosis on radiologist report-drafting time. Journal of Medical Internet Research. 2026;28:e77967. Open source
  19. Healthcare Innovation. Northwestern Medicine Develops Own Generative AI Tool for Radiology. June 10, 2025. Open source
  20. Radiology Business. Radiology AI vendors shift focus to workflow integration and enterprise value. 2026. Open source
  21. Signify Research. Medical Imaging VCs: $3.6bn in a Quickly Maturing Market. Premium Insight. Open source
  22. Thompson A. Top 5 predictions for the imaging IT and AI markets in 2025. Signify Research via AuntMinnie. January 2025. Open source
  23. Radiology Business. Best of the Week: RadNet acquires AI firm for $270M. Newsletter. March 7, 2026. Open source
  24. Radiology Business. Rad Partners buys AI firm for $80M. Newsletter. November 22, 2025. Open source
  25. RadNet, Inc. RadNet Reports Third Quarter Financial Results. Company release. November 8, 2026. Open source
  26. Radiology Business. RadNet confident payers will cover its breast imaging AI solution by 2026. Open source
  27. RadNet, Inc. Second quarter 2025 earnings call transcript. Insider Monkey. Open source
  28. The Imaging Wire. AI Finds Overlooked Cancers in Radiology Reports. November 2026. Open source
  29. The Return on Investment of Artificial Intelligence in Radiology. HealthManagement.org. 2026;26(2). Open source
  30. Radiology Business. FDA clears dozens of new radiology AI tools. Newsletter. December 27, 2025. Open source
  31. AuntMinnie. Home page headlines, November 13 to 16, 2026 (TestDynamics funding; $7.8M jury award). Open source
  32. PolicyRisk. Health Tech Investment Act, S. 1399 and H.R. 6197, 119th Congress. Bill tracker. Open source
  33. Impact of AI-Triage on Radiologist Report Turnaround Time. Preprint summary with turnaround values. October 2025. Open source
  34. Ventra Health. AI in Radiology: Reimbursement Landscape 2025. Open source

Suggested citation: Emrick, K. (2026). The algorithm works. The math does not: An investigation into radiology AI’s return on investment problem. Investigative Research Report, Radiology Leadership Series.

Who Really Profits From Radiology AI? Not the Hospitals Paying for It.

Site directory

Every resource, organized by practice.

Homekellyemrick.com