How do AI triage algorithms work in healthcare, and are they safe for patient care?

AI triage algorithms in healthcare are software systems that sort, prioritize, and route patients based on clinical data — symptoms, vitals, imaging, history — rather than relying solely on human judgment. They are already embedded in radiology departments, emergency rooms, mental health intake lines, and telehealth front doors across the United States. Whether they are safe depends far less on the technology itself and far more on how organizations deploy, monitor, and govern it. The record so far is mixed: Aidoc's radiology algorithms run in more than 900 hospitals and imaging centers and have demonstrably sped up detection of conditions like intra-abdominal free gas and incidental pulmonary embolism, while Kaiser Permanente's use of an algorithm — rather than clinicians — to triage mental health patients has drawn a formal union complaint alleging violations of California law. Understanding both sides of that record is essential for any clinic, health system, or care network considering these tools.

What AI Triage Algorithms Actually Do

Also worth reading: What Is the True ROI of Remote Patient Monitoring for Modern Healthcare Networks in 2026? · How Can a Healthcare AI Governance Strategy Scale Secure Care Coordination? · What Should a Healthcare SaaS TCO Calculator Reveal for Clinics and Care Networks?

At their core, triage algorithms are classification and prioritization engines. They ingest structured inputs — a patient's reported symptoms, vital signs, lab values, imaging studies, or answers to an intake questionnaire — and produce an output: a priority score, a recommended care setting, or a routing decision. In radiology, an algorithm might flag a chest CT for immediate radiologist review because it detected features consistent with a pulmonary embolism. In a primary care call center, an algorithm might decide whether a caller with chest pain is routed to the emergency department, a same-day appointment, or a nurse callback.

The distinction between decision support and decision-making is where much of the controversy lives. A tool that surfaces a finding for a clinician to confirm is fundamentally different from a tool that itself determines whether a patient gets a timely appointment. Kaiser's mental health triage system became the subject of union complaints precisely because, according to reporting from CalMatters and Capital & Main, the algorithm — not a therapist — was making the initial determination of how quickly patients would be seen. Critics argued this functioned as a gatekeeping mechanism that delayed care for some patients, potentially violating California's mental health parity and timely-access requirements. The technology is the same in kind; the governance around it is what differs.

Where AI Triage Is Working Today

The strongest evidence for AI triage sits in medical imaging. Aidoc, one of the most widely deployed vendors, offers algorithms for time-sensitive findings including intracranial hemorrhage, pulmonary embolism, and intra-abdominal free gas, and its tools are in active use at more than 900 hospitals and imaging centers. In these settings, the algorithm does not diagnose — it reorders the radiologist's worklist so that a scan showing a life-threatening finding moves to the top of the queue. This is a narrow, well-bounded task with a clear ground truth, which is exactly the kind of problem where current AI performs reliably.

Similar logic applies to specialty triage benchmarking. Research published in npj Digital Medicine has advanced the idea of benchmarking and competition for specialty triage algorithms, allowing models to be compared head-to-head on standardized datasets. This matters because it moves the field away from vendor marketing claims and toward measurable performance. For care networks evaluating tools, the lesson from imaging is clear: AI triage works best when the task is narrow, the data is structured, the output is a prioritization rather than a final decision, and a human remains in the loop to confirm or override.

Where AI Triage Has Gone Wrong

The mental health cases are the cautionary tale. Reporting by CalMatters, Capital & Main, and San Jose Inside documented allegations that Kaiser Permanente used an algorithm to triage patients seeking mental health care, with the National Union of Healthcare Workers arguing that the system routed patients to lower levels of care than clinically warranted and contributed to long wait times for therapy. The union's complaint framed this as a violation of California law governing access to mental health services. Kaiser disputed the characterization, but the episode illustrates a structural risk: when an algorithm controls the front door to care, any bias or miscalibration in that algorithm scales across the entire patient population.

Health Affairs has explored this problem under the framing of "when the front door is an algorithm," noting that consumers often cannot tell when they are interacting with AI versus a human, and that regulators have been slow to establish disclosure and oversight requirements for these intake systems. The core failure mode is not that the algorithm is malicious — it is that utilization management incentives and clinical priorities can get baked into a model's training data, and patients have no way to appeal a decision they do not know was made by software.

Algorithmic Bias and the Transparency Problem

Algorithmic bias in triage is not hypothetical. Models trained on historical utilization data can reproduce historical inequities: if a population historically underused a service, the model may learn to deprioritize similar patients. The complexity of certain algorithms also poses a genuine barrier to understanding their functioning — many are effectively black boxes even to the organizations deploying them. Worse, algorithms may change over time or respond differently to shifting inputs, meaning a tool validated at launch may drift in performance as patient populations, coding practices, or care patterns evolve.

For a clinic or care network, this translates into concrete due-diligence requirements. Ask vendors whether the model's logic is explainable to a clinician. Ask whether the model has been validated on a population that resembles yours — a triage tool trained on a large urban academic hospital's data may perform differently in a rural clinic. Ask whether the vendor notifies customers when the model is updated, and whether you can run a shadow period where the algorithm's recommendations are compared against clinician decisions before it goes live. A vendor who cannot answer these questions plainly is telling you something important.

Human-in-the-Loop vs. Algorithm-Led Triage: A Comparison

The central design decision any organization faces is how much authority to give the algorithm. The two dominant models look like this:

Neither column is universally correct. A high-volume radiology department gains enormous value from AI worklist prioritization with zero loss of clinician authority. A mental health intake line that lets an algorithm decide who sees a therapist — as the Kaiser allegations describe — carries legal, ethical, and reputational risk that most organizations should not accept. The pattern across documented failures is consistent: risk rises sharply when the algorithm's output substitutes for, rather than informs, clinical judgment, and when the triage decision affects access to scarce services like therapy appointments.

Practical Steps for Clinics and Care Networks Evaluating AI Triage

Organizations considering these tools should treat deployment as a phased clinical project, not an IT purchase. Start by defining the specific triage decision you want to improve and measuring current performance: average time to appointment, percentage of patients routed to the correct care setting, and no-show or escalation rates. Without a baseline, you cannot tell whether the algorithm helped.

Second, run the algorithm in shadow mode for 60 to 90 days, comparing its recommendations against actual clinician decisions on the same patients. Disagreement rates above roughly 10 to 15 percent on high-acuity cases are a red flag. Third, define explicit override rights: any clinician interacting with the system should be able to override it without justification paperwork, and overrides should be logged and reviewed monthly to detect drift. Fourth, disclose to patients when AI is involved in their intake — Health Affairs reporting suggests disclosure is becoming an expectation, and several states are moving toward requiring it. Fifth, contractually require the vendor to notify you of model updates and to support periodic revalidation. Finally, assign named clinical ownership of the algorithm inside your organization. A triage model with no accountable clinician is a liability waiting to surface.

Common Mistakes Organizations Make

The most frequent error is buying a triage algorithm to solve a staffing problem. If your clinic cannot answer patient calls within a clinically appropriate window, an algorithm that sorts the backlog faster does not create clinical capacity — it just rationing it more efficiently, which is precisely the dynamic alleged in the Kaiser mental health case. Algorithms can optimize queues; they cannot create therapists.

The second mistake is skipping validation on your own population. A vendor's published accuracy figures come from their training data, not your patient mix. The third is treating the algorithm as set-and-forget infrastructure. Models drift, referral patterns shift, and a tool that was 95 percent concordant with clinicians at launch can degrade quietly. The fourth is failing to plan for the patient-facing narrative: when patients discover — often through a news report or a union complaint — that software was deciding how quickly they could be seen, the trust damage exceeds anything the efficiency gains offset. The fifth is ignoring state law. California's timely-access standards for mental health care, and similar rules in other states, apply regardless of whether a human or an algorithm made the routing decision.

When to Act, and What It Costs

The timing calculus differs by use case. For imaging prioritization and operational routing (deciding which of two appropriate appointment slots to offer), the technology is mature and the ROI case is straightforward; organizations with meaningful imaging or intake volume should be piloting now. For high-stakes access decisions — mental health, chronic disease management, anything where delay causes harm — the wiser course in 2026 is to deploy AI as a monitoring and escalation layer (flagging patients whose condition is deteriorating between visits) rather than as the gatekeeper to care.

On cost, enterprise radiology AI platforms like Aidoc typically operate on per-study or annual site-license pricing, generally running from the low tens of thousands of dollars per year for a single algorithm at a mid-size facility to well over six figures for multi-algorithm deployments across a health system. Patient-facing triage and care-coordination platforms, including patient-pulse and outreach tools used by clinics and care networks, commonly range from a few hundred to a few thousand dollars per month depending on patient volume and integration depth. The larger cost is organizational: shadow-mode validation, clinician training, and ongoing monitoring typically consume 100 to 200 staff hours in the first year. Budget for that honestly, because organizations that skip it are the ones that end up in the news.

The Bottom Line

AI triage algorithms are neither a solution nor a scam — they are a powerful prioritization technology whose safety is determined by governance. The evidence supports narrow, well-bounded, human-supervised deployments in imaging and operational routing. The evidence warns against algorithm-led gatekeeping for access to scarce clinical services, particularly mental health, where the Kaiser allegations and the surrounding regulatory scrutiny should give any care network pause. The question to ask is never "does AI triage work?" but "who is accountable when it is wrong, and would we be comfortable if our patients knew exactly how it decided?"