What Clinical Decision Support Should and Shouldn't Decide

Why clinicians tune out their alerts

Ask a clinician what they think of their decision support tools and you will often get a version of the same answer: they clicked past the last three alerts without reading them. That reflex is not carelessness. It is a rational response to years of being interrupted by software that cries wolf, and it is the central problem any behavioral health AI has to solve before it can help anyone.

The instinct in the industry is to treat that problem as a signal to build smarter models that decide more. The more useful move is the opposite. The question worth answering is not whether AI can make the call, but where it should hand the call back to the person accountable for the patient. Draw that line well and decision support becomes something clinicians reach for. Draw it poorly and it becomes one more thing they route around.

The evidence: even the most severe alerts get overridden

Clinicians do not ignore decision support because they distrust technology in the abstract. They ignore it because too much of it has been wrong, or irrelevant, too often. The numbers are stark. In one evaluation of a drug-interaction clinical decision support system, prescribers overrode 88.2% of the most severe alerts the system produced. These were not the low-stakes reminders. They were the alerts flagged as most serious, and they were dismissed the vast majority of the time.

An override rate that high is not a story about reckless clinicians. It is a story about a tool that lost the room. When a system interrupts often enough with warnings that do not change the decision, clinicians stop reading the warnings, and they stop reading them uniformly, including the ones that would have mattered. The support has effectively deactivated itself while continuing to fire.

Two forces behind alert fatigue: overreach and volume

Two forces drive that erosion, and they compound each other:

1. Overreach: a tool that tries to decide too much, second-guessing judgment calls that belong to the clinician, quickly reads as noise rather than help.
2. Volume: alert fatigue is a well-documented phenomenon, and it scales with load. Research on a decision support system found that a clinician’s acceptance of reminders dropped by about 30% with each additional alert in a patient encounter, and dropped further as repeated alerts piled up.

Every extra interruption does not just fail to help. It measurably lowers the odds that the next alert, including a genuinely important one, gets acted on.

Put those two forces together and the failure mode is clear. A system that decides too much, too often, trains the people it is meant to support to ignore it. The tool that cries wolf gets tuned out precisely when the wolf is real. That is not a reason to abandon decision support. It is a reason to be far more deliberate about what the support is for.

A second set of eyes, not a second opinion that overrides

The framing that resolves this is old and clinical: a good second set of eyes. A trusted colleague who reviews a case does not seize the decision. They notice what you might have missed, name it, and hand the judgment back to you. The value is in the noticing, not in overruling. That is exactly the posture responsible AI should take in a clinical setting.

Applied to behavioral health, that means AI earns its keep by surfacing signal a busy clinician does not have the bandwidth to catch on their own. Analyzing a patient’s speech, expression, and behavioral patterns across assessments and between visits, the model can surface the shifts worth a closer look. What it does not do is convert those flags into a diagnosis or a treatment change. It brings the pattern forward. The clinician decides what it means and what to do about it. The human stays in the loop because the human is accountable for the outcome, and no model carries that accountability.

This is also what protects the trust the first two sections described as fragile. A tool built to surface rather than to decide has a narrower, cleaner job, and it can be tuned to be selective about when it speaks. Selectivity is what keeps it credible. For a concrete example in a shipped tool, see how Videra’s TDScreen shows its work: it surfaces the evidence behind every finding and leaves the decision to the clinician.

Where the line belongs

The practical version of this is a division of labor, drawn on purpose. On the AI side of the line sits the work that benefits from tireless, consistent attention across large volumes of data: surfacing patterns, triaging who may need attention first, and flagging changes that warrant a human look. These are tasks where software genuinely outperforms an overloaded clinician, and where a missed signal is the real risk.

On the clinician side sits the work that requires clinical judgment, context, and accountability: diagnosis, treatment decisions, and the interpretation of what a given signal means for a particular person with a particular history. AI can inform these. It should not make them. The line is not drawn by what a model is technically capable of attempting. It is drawn by where a human must remain responsible, and in clinical care that line sits firmly at diagnosis and treatment.

Drawing it this way is not a limitation grudgingly accepted. It is the design principle that makes the whole system safe to use. When clinicians know the tool will not try to overrule them, they can let it into their workflow without bracing against it.

What good looks like

Decision support built on this principle behaves differently from the tools that trained clinicians to click past alerts. It is quiet more often, because it only speaks when the signal justifies it. When it does speak, the flag is specific and traceable back to the underlying evidence, so a clinician can evaluate it in seconds rather than dismiss it on reflex. In practice that looks less like a pop-up warning and more like a specific, reviewable note, for example a widening gap between what a patient reports and how their speech and expression have presented across recent check-ins, surfaced with the underlying responses attached so the clinician can judge it in context. And it never removes the clinician from the decision, which means adopting it does not require surrendering control.

The result is a tool that gets used, and one that makes care safer for a reason that has nothing to do with the sophistication of the model. It is safer because the human stays in the loop, the signal reaches them before it would have surfaced on its own, and the judgment that acts on that signal belongs to someone who can be held accountable for it. That combination, an alert worth reading plus a clinician still firmly in charge, is the whole point.

Press vendors on where the AI stops

The behavioral health organizations that get the most out of AI will not be the ones that let it decide the most. They will be the ones that are clearest about where it should stop. That clarity is what turns decision support from a compliance checkbox into something clinicians actually rely on, and it is the question worth pressing any vendor on before a deployment. If a tool cannot tell you where it hands the decision back, that is the answer.

For a fuller view of how behavioral health systems are drawing these lines as they move AI from pilot to daily use, Videra Health’s research report The State of AI in Behavioral Health 2026 lays out where the value is showing up and what clinician acceptance actually looks like in practice. And if you want to see where Videra draws the line in a live workflow, book a demo.