From Prediction to Navigation: The AI Shift Healthcare Has Been Waiting For

For years, the most familiar promise of artificial intelligence in medicine has sounded almost prophetic. An algorithm sees what people miss. It detects a tumor hiding in a scan, predicts who may crash overnight, flags a patient likely to be readmitted, or finds a pattern buried too deeply in the record for a clinician to catch during a 17-minute visit.

Presented By Our Partners

That kind of prediction matters. But prediction is not care.

A risk score does not schedule the follow-up colonoscopy. A sepsis alert does not decide whether a patient needs fluids, vasopressors, antibiotics, ICU transfer, or simply a better threshold for alarm. A breast-cancer screening model does not make sure the woman it flags actually gets diagnostic imaging, biopsy, treatment planning, transportation, coverage approval, and a human being who owns the next step.

This is the quiet gap in medical AI. Healthcare has built an increasingly powerful signal machine. It has not yet built an equally powerful action machine.

That distinction is now becoming harder to ignore. In a recent Lancet comment, Girish Nadkarni and colleagues argue that clinical AI needs to move “from prediction to navigation.” The phrase is useful because it names the missing layer. Most clinical AI systems still estimate what is present or what might happen. Navigational AI would ask the more difficult question: what should happen next, for this person, in this context, at this moment?

That is not a cosmetic upgrade. It is the difference between an alarm and a route.

The problem with medicine’s signal era

The first wave of clinical AI was built around detection and prediction. That made sense. Medicine is full of pattern-recognition problems, and many of them are cognitively exhausting. Radiologists read huge image volumes. Hospital teams monitor dozens of unstable patients. Primary-care clinicians try to identify the one dangerous presentation hiding inside a day packed with ordinary complaints.

But the prediction era has produced a familiar disappointment: technically impressive models that do not reliably change outcomes.

Nadkarni and colleagues point to a widely implemented sepsis prediction model whose positive predictive value was only 2.4% at a commonly used alert threshold for predicting sepsis within 24 hours. Put plainly, more than 97% of alerts were false positives in that window, while the model still missed more than 67% of patients who developed sepsis. That is not just a statistics problem. It is a workflow problem. False alarms do not merely fail to help; they consume attention, train clinicians to distrust the system, and add noise to an environment already saturated with alerts.

This is the core challenge for healthcare delivery. A prediction can be accurate enough to impress in a paper and still be too blunt to improve care. If the output does not help someone choose the next action, assign responsibility, reduce friction, or complete a care pathway, it may widen the distance between clinical knowledge and clinical reality.

That distance is the Discovery-to-Delivery Gap: the space between knowing something useful and making it happen for the patient who needs it.

Navigation means choosing, not just noticing

Navigation is harder than prediction because medicine is not a single-turn game. A clinician does not merely ask, “Is this patient at risk?” The real questions come in sequence.

Should this patient go to the emergency department or primary care? Should the next visit be urgent, routine, virtual, or unnecessary? Is this abnormal scan finding a low-risk incidentaloma or the start of a cancer workup? Should a patient with sepsis receive more fluid or less? Which follow-up step is clinically appropriate, covered by insurance, available nearby, and likely to be completed?

Each decision depends on physiology, history, access, patient preference, coverage rules, local capacity, and time. A useful AI system has to understand more than the signal. It has to understand the path.

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 →

That is why the strongest near-term story in healthcare AI may not be the most futuristic one. It may be the software that moves patients through boring, fragile, failure-prone steps: symptom assessment, triage, scheduling, prior authorization, referral completion, follow-up tracking, and escalation when the system goes quiet.

Healthcare often fails in these handoffs. The suspicious imaging finding is noted but not acted on. The referral is placed but never completed. The specialist visit is delayed by insurance friction. The patient is told to follow up, but no one owns the loop. The clinical signal exists. The delivery system drops it.

A real-world hint from the digital front door

A March 2026 NEJM AI study offers a useful early example of what “signal to action” can look like outside the hospital ward. The ESSENCE study evaluated an AI-supported symptom assessment system embedded in the myCUF app inside Portugal’s largest private healthcare network.

This was not just another accuracy study. The investigators looked at whether the tool changed what people intended to do and what they actually did afterward.

Among 1,470 adults, 33.0% of participants with pre- and post-assessment intentions revised their planned care level immediately after using the system. Uncertainty fell from 12.6% to 5.0%. Among 721 participants with observed behavior, 59.1% changed their care pathway: 28.9% de-escalated, 17.2% escalated, and 13.0% resolved prior uncertainty. Primary-care consultations rose from 16.3% to 42.1%, while specialist visits fell from 49.7% to 29.8%.

The most important number may be the one closest to actual care quality. Among nonemergency participants with sufficient clinical documentation, physician-rated appropriate care increased from 29.8% before assessment to 64.4% after observed behavior.

This does not prove that every digital front door should be treated as a clinical navigator. The study was conducted in one network, with a specific tool, in a specific context, and the authors call for further evaluation across diverse settings. But it does show the right measurement instinct. The question was not only whether AI produced a plausible recommendation. The question was whether patients moved toward more appropriate care.

That is the bar AI in healthcare should be asked to clear more often.

The payer side cuts both ways

Navigation can improve access, but it can also become a more efficient gate. Nowhere is that tension clearer than prior authorization and claims review.

A recent KFF issue brief describes how AI is being adopted across the coverage-review cycle by insurers, providers, and patients. Insurers use automation and AI to process claims, triage coverage decisions, and support utilization management. Providers use AI inside revenue-cycle workflows to assemble documentation, check eligibility, code visits, and accelerate prior authorization. Patients and clinicians are beginning to use AI tools to generate appeal letters and denial-response documentation.

The same technical layer can therefore push in opposite directions. It can help a patient get approved faster for necessary care. It can help a health system reduce administrative drag. Or it can help a payer deny care at scale if governance is weak, transparency is poor, or human review becomes ceremonial.

KFF cites a National Association of Insurance Commissioners survey in which 84% of responding insurers across health insurance product lines reported using AI or machine learning for a broad range of tasks, including utilization management, disease management, and prior authorization. That is not a distant future. It is already operational.

This is why “navigation” cannot simply mean “AI tells people where to go.” A navigation system in healthcare must be judged by whether it improves appropriate access, reduces avoidable delay, preserves appeal rights, and makes responsibility visible. Otherwise, healthcare may replace one maze with another — faster, shinier, and harder to contest.

The new test: completed care

The next phase of medical AI should be evaluated less like a magic detector and more like an operating system for follow-through.

Did the patient receive the right care setting? Did the abnormal result trigger a completed diagnostic pathway? Did the referral close? Did the prior authorization clear without avoidable delay? Did the patient understand the next step? Did the system reduce inequity, or did it simply navigate the well-resourced faster?

These are less glamorous questions than whether a model can outperform a physician on a benchmark. They are also closer to what people actually need from healthcare.

This is where the best healthcare AI companies and health systems may separate themselves. Case finding is one layer: identifying patients who medicine is already missing. Navigation is the next: moving those patients through the operational maze. Closed-loop delivery is the final test: proving that the care was completed and that outcomes improved.

That is the arc from signal to action to completed care.

The future is not an AI doctor. It is an accountable route.

The seductive version of medical AI imagines a machine that knows the answer. The more useful version may be a system that helps medicine stop losing the thread.

A patient has a symptom. A scan has a finding. A lab result changes. A risk score rises. A prior authorization stalls. A referral sits untouched. A treatment exists but never reaches the person who qualifies for it.

The important question is not whether AI can notice more of these signals. It can. The question is whether AI can help healthcare carry the signal all the way to action.

Prediction made AI visible in medicine. Navigation will determine whether it becomes useful.

And completed care — not clever alerts, not elegant demos, not benchmark victories — is where the judgment should land.

Sources

• Nadkarni GN, McGreevy JD, Yu KH, Singh K, Sakhuja A. “From prediction to navigation for artificial intelligence in medicine.” The Lancet. Published online April 27, 2026. https://doi.org/10.1016/S0140-6736(26)00756-7
• Cotte F, Dias Lourenço F, Paiva Pereira M, et al. “From Advice to Action — Real-World Behavior of Patients Using an Integrated Diagnostic Decision Support System for Navigating the Health Care System.” NEJM AI. Published March 5, 2026. https://doi.org/10.1056/AIoa2500833
• KFF. “Regulation of AI in Prior Authorization and Claims Review: A Look at Federal and State Consumer Protections.” May 2026. https://www.kff.org/patient-consumer-protections/regulation-of-ai-in-prior-authorization-and-claims-review-a-look-at-federal-and-state-consumer-protections/

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.