Medicine Has Enough AI Pilots. It Needs a Power Grid.

A hospital can now buy an artificial-intelligence tool for almost any fragile point in care.

Presented By Our Partners

One system reads a CT scan and flags a suspicious lung nodule. Another combs through a thousand-page oncology record and suggests the next trial a patient might qualify for. Another gathers the documentation needed for prior authorization. Another listens in the exam room and turns a visit into a note. Another watches operational data and predicts who may need a post-acute bed.

This is the strange new abundance of medical AI: the pieces are arriving faster than the system that would make them useful.

The problem is no longer simply whether an algorithm can notice something. It is whether anyone owns what happens after the noticing.

A recent Nature commentary put a sharper name on the missing layer. The authors argued that the bottleneck in AI-enabled health systems was “never data or algorithms.” What medicine needs, they wrote, is a “learning utility”: permanent infrastructure that connects data, AI, workflow, action, outcomes, monitoring, and governance.

That phrase is technical, but the metaphor is plain. Modern life does not run because every house owns a generator. It runs because electricity became a grid. Healthcare AI is still mostly in the generator era.

Hospitals are buying and building impressive local machines. Some are genuinely useful. Some are overhyped. Many are somewhere in between. But too often each tool lives as a project: a pilot, a vendor integration, a dashboard, a committee-approved deployment. The signal may be real. The next step may be unclear. The outcome may never be captured. The model may never learn whether it helped.

That is not a small operational detail. It is the difference between AI that decorates healthcare and AI that changes healthcare delivery.

Healthcare delivery is the chain between what medicine knows and what a person actually receives: signal, ownership, action, completion, and measured result. AI can strengthen that chain. It can also make the chain noisier if the health system treats prediction as the finish line.

The next serious test for medical AI is not whether it can generate more alerts, summaries, matches, or recommendations. The test is whether it can close the loop.

The Signal Is Getting Easier

Consider lung cancer.

A collaboration announced last week by Massive Bio and Optellum is a clean example of where the field is trying to go. Optellum’s platform analyzes radiology findings and CT cases to surface patients at elevated risk for lung cancer before diagnosis is fully confirmed. Massive Bio’s platform then matches de-identified patient profiles to geographically accessible, active non-small cell lung cancer trials and sends actionable reports to treating physicians.

In the companies’ framing, the ambition is not merely to detect risk. It is to connect a high-risk CT finding with a trial referral while the window for curative-intent treatment may still matter.

That is the right shape of the problem.

A suspicious image is not care. A risk score is not care. A trial match is not care. Care begins to emerge only when a responsible clinician receives the signal, understands the patient context, discusses options, navigates eligibility and logistics, and gets the patient to the next appropriate step.

Oncology shows the same pressure from another direction. Triomics, an AI oncology company profiled by Forbes, says its platform processes structured and unstructured cancer records to help clinicians identify next treatment steps and potential trials. The company has reported a 67% reduction in chart-review time and a 40% improvement in trial matching, with customers including major cancer centers and community oncology practices.

Those numbers should be read with appropriate caution. Company-reported and conference-linked performance claims are not the same as independent proof of improved survival, quality of life, or equitable access. But the workflow target is important. In cancer care, delay often hides inside paperwork: scans, pathology, genomic reports, referral notes, prior treatments, eligibility criteria, and payer requirements scattered across systems that do not behave like one system.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

AI may be useful here because the human bottleneck is not lack of expertise. It is the impossible information load between expertise and action.

The Paperwork Is Also Part of Care

It is tempting to think of prior authorization, documentation, benefits checks, and referral tracking as administrative problems adjacent to medicine. Patients experience them differently. A denied authorization, a missing form, a forgotten referral, or an unreturned call can be the practical end of care.

That is why prior authorization automation belongs in the same conversation as diagnostic AI. HealthTech Magazine recently described prior-auth tools that gather clinical data from the EHR, check payer requirements, flag missing information, fill out forms, and track request status. The promise is familiar: instead of waiting days or weeks, some requests may be processed in minutes.

The danger is also familiar. If automation simply accelerates a broken process, it can make denial, opacity, and bureaucratic burden faster. But if the tool is connected to accountability — what was requested, what was missing, who followed up, whether the patient received the service — then it becomes part of the care pathway.

This is where the “learning utility” idea matters. It asks health systems to stop treating AI as a set of clever point solutions and start treating it as institutional infrastructure. A useful AI deployment should know what signal it generated, who received it, what action was taken, what outcome followed, and whether the system should keep using the tool in the same way.

Without that return path, medicine cannot reliably answer the most important question: did this help the patient get care?

Governance Has to Move After Go-Live

Healthcare has historically treated safety evaluation as something that happens before a tool goes live. That is necessary, but it is no longer sufficient.

AI systems can drift. Patient populations differ. Workflow incentives vary. A model that performs acceptably in one hospital may fail quietly in another. A generative system that seems helpful in a controlled setting may behave differently when connected to real documentation, real patients, and real time pressure.

The governance conversation is beginning to catch up. The Joint Commission’s new Responsible Use of AI in Healthcare certification, highlighted in its June 2026 news, focuses on organization-level governance rather than certifying individual products. Its domains include governance, data management, risk and bias reduction, monitoring and validation, transparency, education, and training.

Mount Sinai researchers also published work this month describing the fast-growing but fragmented policy landscape around healthcare AI. Their point was not that one rule will solve everything. It was that health systems are now operating in a patchwork of regulations, standards, institutional guidance, and technical expectations.

That patchwork creates a practical burden. Somebody has to translate policy into workflow. Somebody has to decide what gets monitored. Somebody has to notice when an AI tool is no longer helping. Somebody has to de-implement tools that fail.

The Nature commentary makes this point with unusual clarity: the missing layer is not just software. It is people and roles. Clinical-technical translators. Research-operations translators. Governance-technical translators. Care-patient translators. The grid needs linemen, engineers, inspectors, and dispatchers. Healthcare AI needs its own version of that workforce.

The New Bar: From Prediction to Completion

For the past decade, many AI-in-healthcare debates have centered on whether algorithms could match expert performance. Could a model read an image? Predict a risk? Draft a note? Identify a patient? Summarize a record?

Those questions still matter. But they are no longer enough.

The better question is what happens next.

If AI flags a patient at high risk for lung cancer, is there a pathway to diagnostic confirmation, specialist review, trial matching, treatment, and follow-up?

If AI identifies a patient eligible for a clinical trial, does anyone contact the patient, check geography and eligibility, coordinate with the treating oncologist, and measure whether enrollment actually happened?

If AI speeds prior authorization, does the patient receive the scan, medication, procedure, or referral? Or does the system merely move the bottleneck somewhere else?

If AI drafts a perfect note, does it free a clinician to listen better, close a care gap, or catch a missed next step? Or does it simply create a more polished record of the same fragmented care?

These are not anti-AI questions. They are pro-outcome questions.

The most important healthcare AI systems of the next few years may not look spectacular in a demo. They may look like boring infrastructure: registries, feedback loops, monitoring dashboards, referral workflows, payer checks, outcome capture, de-implementation protocols, and human escalation paths.

That is usually how serious progress looks once the novelty wears off. The miracle becomes plumbing.

The Real Promise Is Less Abandonment

Patients do not usually fall through the cracks because medicine lacks information. They fall through because information does not reliably become action.

A scan is abnormal, but follow-up is delayed. A trial exists, but no one matches the patient in time. A treatment is appropriate, but authorization stalls. A social need is documented, but the referral never closes. A risk score fires, but ownership is ambiguous. A dashboard glows, but no one has the staffing, authority, or workflow to act.

AI can worsen that problem by producing more signals than the system can absorb.

Or it can help solve it by making every signal accountable to an action and every action accountable to an outcome.

That is the future worth watching. Not artificial intelligence as a pile of dazzling tools, but AI as part of a learning healthcare utility: connected, monitored, governed, and judged by whether people actually receive the care medicine already knows how to provide.

The next breakthrough may not be the model that sees the problem first.

It may be the system that makes sure the problem does not disappear after being seen.

Sources

• Nature, “The bottleneck was never data or algorithms: building a learning utility for AI-enabled learning health systems,” June 2026.
• BioSpace / company announcement, “Massive Bio and Optellum Collaborate to Accelerate Clinical Trial Access for Early-Stage Non-Small Cell Lung Cancer Patients,” June 2026.
• Forbes, “How Triomics Uses AI To Help Doctors Treat Cancer More Quickly,” June 2026.
• HealthTech Magazine, “Clinical Workflow Automation: Where AI Is Making Real Inroads in Healthcare,” June 2026.
• The Joint Commission, June 2026 Responsible Use of AI in Healthcare certification update.
• EurekAlert / Mount Sinai, “Researchers create first-of-its-kind index of evolving policy landscape around health care AI,” June 2026.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.