Clinical AI in 2026: three deployment tiers operators should plan around

If you are a COO or CMIO at a regional hospital group or a national chain, the question on your desk in 2026 has stopped being whether to do clinical AI. The live question is which clinical AI to fund this fiscal year, which to pilot, and which to defer to the 2027-2028 capex cycle. Most vendor decks treat all of that as one decision. The evidence does not.

By May 2026 the FDA has authorised roughly 1,451 AI-enabled medical devices since it began counting in 1995, and about 76 percent of them are in radiology (per The Imaging Wire, March 2026, citing FDA's running list). That figure is what makes the bedside-vs-research framing wrong. A large body of clinical AI is cleared, paid for, and used in some health systems every shift, while the same product sits unused in others. The interesting cut in 2026 sits between three tiers rather than between deployed and research: clinical-default, cleared-but-uneven, and still-research-grade.

This piece sets out those three tiers, what evidence supports each, what binds under the EU AI Act on 2 August 2026, and the procurement floor every tier shares. It is written for the operator who has to defend a number in a board paper, not for the bench scientist and not for the editor of a vendor briefing.

The bedside-vs-research split is the wrong frame

The reason the dichotomy fails is that the gap is not capability. It is workflow integration and local validation. PathAI's AISight Dx platform has been FDA-cleared for primary diagnosis since June 2025, and got a second clearance (AISight Dx2) with a Predetermined Change Control Plan in March 2026 (per PathAI's announcement). Yet fewer than ten US health systems run primary-diagnosis whole-slide AI at scale. The clearance is real. The deployment is not. That is the pattern across most "transformative" clinical AI in 2026: cleared, available, paid-for in some markets, and barely visible in others.

The reason matters for budgeting. If a category is cleared-but-uneven, the procurement question is "what is the local validation cost and the integration timeline," not "does this technology work in principle." If a category is still-research, the procurement question is "what should the watch-budget be." Conflating the two is how mid-market health systems end up writing eight-figure cheques for technologies their workflows cannot consume.

Tier 1: deployment-default in 2026

Two categories cleared this bar by May 2026.

Radiology triage AI. Aidoc received FDA clearance in January 2026 for the first comprehensive AI triage solution, with 97 percent mean sensitivity and 98 percent mean specificity across 11 indications in the pivotal study (per PRNewswire, January 2026). For large vessel occlusion stroke specifically, Aidoc reported 92.6 percent sensitivity against 70.4 percent for conventional solutions. Viz.ai's stroke alert data, published in multiple peer-reviewed studies, shows patients reach treatment around 66 minutes faster when the AI alert system is active. Radiology owns roughly 1,104 of the 1,451 cleared AI devices because the modality is well-suited to image classification, the workflow already routes through PACS, and the value proposition (faster triage of acute findings) maps directly to a billable outcome.

Ambient AI scribes. A pragmatic randomised trial published in NEJM AI in 2025 covered 238 outpatient physicians across 14 specialties, randomised to Microsoft DAX Copilot, Nabla, or usual care (per the NEJM AI publication). Both ambient tools reduced documentation time and showed signals of burnout reduction. A separate UCSF study in JAMA Network Open (September 2025) found physicians using AI scribes generated 3,044 dollars more revenue per year and saw 0.8 additional patients per week. A second 2025 JAMA paper across two large health systems reported a 21 percent burnout reduction. The evidence is RCT-grade, replicated across systems and vendors, and the integration cost is modest because the input is audio and the output is structured EHR text.

If your board has not approved spend on either of these two categories yet, that is the easier paper to write in 2026. The harder question is what comes next.

Tier 2: cleared, uneven adoption

This is where most of the 2026 procurement decisions actually live.

Pathology AI. Around 51 FDA-authorised pathology AI devices through April 2026, of which only seven are whole-slide imaging algorithms. PathAI's AISight Dx and Paige's PanCancer Detect (FDA Breakthrough Device designation, April 2025) are technically capable; the bottleneck is that primary-diagnosis whole-slide pathology requires a fully digital workflow, which most US labs still do not run. Labcorp's expanded PathAI roll-out (announced 2025-2026) is one of the few national-scale moves. For mid-market health systems the procurement question is not "is the AI good enough." It is "is our pathology lab already digital, and what is the conversion cost." The answer for most operators is "no" and "seven figures over two years."

Sepsis early-warning systems. This category carries the loudest cautionary tale in clinical AI. The Epic Sepsis Model v1, deployed at hundreds of US hospitals, was externally validated in JAMIA Open (November 2024) across two county emergency departments covering 145,885 encounters: 14.7 percent sensitivity, 95.3 percent specificity, 7.6 percent positive predictive value (per the JAMIA Open paper). A multicentre validation of Epic Sepsis Model v2 across four large US health systems, with data through March 2025, is in late-stage analysis as of mid-2026. The point is not that AI sepsis prediction is useless. The point is that a vendor-trained model's local performance can be far below its development-set numbers, and the only way to find out is to validate it on your own population before turning it on in clinical workflow. A 2025 meta-analysis covering ten RCTs of AI-based ICU decision support across roughly 100,000 patients found mostly improvements in process measures rather than consistent mortality effects.

Wearable AFib detection and remote patient monitoring. The EQUAL trial published in JACC in January 2026 (per the JACC publication) randomised 437 patients aged 65 and over with elevated stroke risk to Apple Watch-based monitoring versus standard care. The smartwatch group showed improved detection of new-onset and asymptomatic AFib. The CMS 2026 Physician Fee Schedule Final Rule lowered RPM time thresholds effective 1 January 2026, which removed one of the main reimbursement frictions. The constraint here is operational: who reads the alerts, who follows up with the patient, and how is that staffed.

AI-supported personalised treatment plans. A Journal of Clinical Oncology paper (8 January 2026 online edition) from UC San Diego found that individualising multi-drug regimens to each tumour's molecular profile, with AI-assisted selection, significantly improved outcomes. The functional precision medicine approach (combining AI with live tumour-cell testing) has reported up to 83 percent outcome improvements over standard care in peer-reviewed work. The categories where this is becoming routine procurement are oncology centres of excellence and a small number of academic medical systems. The mid-market hospital question is whether the population served has enough complex oncology to justify the molecular tumour board infrastructure, not whether the AI works.

Tier 3: still research

Three categories take the loudest vendor airtime and have the weakest deployment evidence in May 2026.

Autonomous or semi-autonomous robotic surgery. Intuitive Surgical's da Vinci 5 received FDA clearance in March 2024 and an expansion clearance for cardiac procedures (including mitral valve repair) in 2026 (per Intuitive's IR release). CMR Surgical's Versius Plus received its FDA clearance in December 2025 with US commercial launch in 2026 (per The Robot Report, December 2025). Both are surgeon-controlled platforms with AI-assisted feedback. There is no FDA approval as of May 2026 for any autonomously-performed surgical procedure. The capex argument for surgical robotics is a real one (Intuitive's installed base passed 10,000 systems years ago) and the argument is about precision, ergonomics, and force feedback rather than autonomous action.

AI-designed therapeutics in human trials. Insilico Medicine's ISM001-055 for idiopathic pulmonary fibrosis posted positive Phase IIa topline results in November 2024 (71 patients, dose-dependent improvement in forced vital capacity at 12 weeks, per the Insilico press release). That is the most-advanced AI-discovered candidate in clinical development as of mid-2026. Isomorphic Labs raised 2.1 billion dollars in May 2026 and now targets first-in-human trials by the end of 2026, slipping from Demis Hassabis's original end-of-2025 target (per TechStartups, May 2026). Recursion's lead candidate REC-994 for cerebral cavernous malformation was discontinued in May 2025 after long-term data did not confirm earlier efficacy trends. Exscientia's lead psychiatric candidate DSP-1181 was discontinued in 2022. BenevolentAI's atopic dermatitis candidate BEN-2293 missed efficacy endpoints and the company was acquired in March 2025. AI changes drug-discovery odds. It does not change drug-development biology. For a hospital operator the implication is that 2026-2030 therapeutic pipelines are still mostly conventional medicinal-chemistry products with AI-assisted optimisation, not AI-designed novel mechanisms.

Generative-AI diagnostic and symptom-checker chatbots. Symptom-checker accuracy studies continue to publish underwhelming numbers: Ada Health around 22 percent, Babylon around 41 percent, Symptomate around 51 percent for orthopaedic pathology in a 2024 multi-observer study. The FDA held a dedicated Digital Health Advisory Committee meeting on generative-AI mental-health devices in November 2025, and Illinois banned AI from making independent therapeutic decisions without licensed professional review in August 2025. Babylon Health collapsed in 2023; Woebot's consumer app shut in 2025. The category that does work is documentation assistance, which we covered under Tier 1 as ambient scribes. Front-door symptom triage, autonomous diagnosis, and generative therapy chat are still research.

What changes on 2 August 2026

The EU AI Act's high-risk obligations under Annex III bind on 2 August 2026, with the exception of AI systems embedded in CE-marked medical devices under Annex I, which slip to 2027 (per artificialintelligenceact.eu). Clinical-decision-support AI, patient-triage systems, and emergency-call-prioritisation AI all sit in Annex III, point 5. The fines are up to 15 million euros or 3 percent of global annual turnover, whichever is higher.

The operator implication is that hospitals using third-party AI tools are deployers under the Act, not just providers, and the deployer obligations bite in less than three months. Audit trails, human oversight, conformity-assessment evidence from the vendor, and post-market monitoring all become hard requirements for tier-1 and tier-2 systems. Our healthcare AI governance pillar covers the deployer-versus-provider split and what evidence the contract has to compel from the vendor. If the procurement contract you are signing in 2026 does not name a conformity assessment body and a CE marking pathway, that is a red flag.

The procurement floor every tier shares

Three things every clinical AI buy in 2026 needs, regardless of tier.

Local validation budget. Real-world deployment of cleared AI shows performance drops of 15 to 30 percent versus benchmark, driven by population shift and integration friction (Stanford-Harvard State of Clinical AI 2026; Frontiers in Medicine 2025). A radiology triage product cleared on a national-level FDA dataset is not automatically calibrated to your patient mix. Either the vendor runs the local validation as part of deployment, or you fund a clinical-informatics team to do it. Skipping this is how a hospital ends up replicating the Epic Sepsis Model story on a different model.

Deployer-side audit infrastructure. The EU AI Act and FDA post-market guidance both move some monitoring responsibility from the vendor to the deployer. Tier 1 and Tier 2 systems need logging, drift detection, and equity-of-performance reviews stratified by patient subgroup. Pulse-oximeter racial bias and the broader 2024 UK government-commissioned review of medical-device bias are the cautionary tales that have already cost regulator attention.

A walk-away clause. Tier 2 procurements especially benefit from a contractual exit if local validation does not replicate the vendor's claims within a defined window. The vendor's clearance is necessary, not sufficient.

What the evidence does not say

This is the paragraph the reader judge cares most about, so let me be specific.

The evidence does not say AI reduces mortality across most clinical categories with statistical reliability. A 2025-2026 systematic review of RCTs in adult ICU AI decision support covering roughly 100,000 patients found mixed mortality results, with most studies improving process measures (time-to-detection, time-to-treatment) rather than clinical end points. The two RCTs that did show mortality reductions were specific to sepsis prediction and machine-learning-based early-warning in particular settings. Generalising from "AI catches X faster" to "AI saves lives" is the move where vendor decks lose evidentiary footing.

The evidence also does not say that the FDA-cleared tier 1 categories work equally well across populations. Pulse oximeter bias against darker skin tones (well-documented in the New England Journal of Medicine and Johns Hopkins work since 2020) is a hardware story; the equivalent algorithmic-bias story is younger but growing, and the FDA has signalled new guidance is forthcoming. Tier 1 should not mean "deploy without monitoring." It means "the clearance evidence is strong enough to deploy with monitoring."

And the evidence does not say that AI-designed therapeutics will fail. The Insilico IPF Phase IIa is a real signal, and the Isomorphic-Labs platform is well-funded and credibly led. It says that as of May 2026 the deployed-medicine reality is that AI shifts drug-discovery probabilities at the front of the pipeline, not at the back. The first wave of AI-designed medicines in human use is plausibly a 2028-2030 story, not a 2026 procurement line.

What you do next

If you are writing the 2026-2027 clinical AI capex paper this quarter, the structure is roughly: fund tier 1 now with local validation budget attached, run two or three tier 2 pilots with walk-away clauses and stratified-performance reporting, and book the tier 3 categories as watch-budget with a quarterly review against the regulatory and clinical-trial calendar.

The two questions that matter more than the tier framing are: which of these categories your patient population actually consumes value from (a stroke-heavy referral catchment monetises tier 1 radiology triage faster than a paediatric-heavy one), and whether your data and informatics infrastructure can ingest the AI's output into a workflow a clinician already trusts. The harder of those two is usually the second.

We build software for healthcare operators and have shipped against this space for several years. RecoverMe is a CBT-based app for gambling-addiction recovery now used at scale across the UK, and the Brighton and Great Western Health community platforms we run are how some NHS services route patient interactions. If you are scoping the workflow-integration side of one of these clinical AI buys and want a build partner who has shipped against real NHS and EU healthcare environments, we are around. If not, the tiering above should still be a defensible spine for your board paper.