Two AI outputs land in front of the same clinician within the same minute. The first is a drafted reply to a patient asking for the clinic's parking instructions. The second is an AI-suggested insulin dose for an inpatient with fluctuating glucose and impaired renal function. If the clinician gives both the same amount of scrutiny, one of two bad things is happening: either she is wasting careful attention on parking instructions, or, far worse, she is spending only a parking-instructions level of attention on an insulin dose. Vigilance is finite. The entire craft of this lesson is learning to spend it where the danger actually lives.

Vigilance Is a Finite Budget

The previous lessons established that a real human-in-the-loop must genuinely verify, and that the handoff must place a sign-off gate before any patient-facing consequence. But those lessons leave a practical question unanswered, and it is the question that decides whether a workflow survives contact with a real shift: how deeply must the human check each output? The naive answer, "check everything equally and thoroughly," is not just impractical; it is actively unsafe, because it ignores the one resource that is genuinely scarce in clinical work. That resource is not time exactly, and it is not effort exactly. It is vigilance: the finite capacity for careful, skeptical attention that a human can bring to bear before it degrades into fatigue and inattention.

Treat vigilance as a budget, because it behaves like one. Every output you scrutinize deeply spends some of it. If you spend the same deep scrutiny on a parking-instructions reply as on an insulin dose, you are not being safe; you are being indiscriminate, and indiscriminate vigilance runs out. By the time the high-stakes output arrives, you have exhausted the very attention it needed, squandered on outputs where a mistake would have been trivial. The clinician who tries to check everything at maximum depth does not achieve maximum safety. They achieve uniform, mediocre attention that is too shallow where it matters and wastefully deep where it does not. The goal is not more vigilance everywhere. It is vigilance matched to stakes, concentrated where a wrong output could actually hurt someone.

An analogy: the headlamp on a long night

Picture vigilance as the battery in a headlamp you must wear for a twelve-hour night shift, with no chance to recharge until morning. You could run it at full brightness the moment you clock in, illuminating every corner of every room with equal intensity. It would feel diligent. It would also be a mistake, because the battery is finite and the darkest, most dangerous passages of the night are still ahead. The disciplined caver does not burn the lamp at maximum on the wide, flat, well-known corridors. She dims it there so that the beam is at full strength when she reaches the narrow ledge above the drop, where a misstep cannot be undone. Risk-based verification is that discipline applied to attention. Dimming the lamp on the parking reply is not carelessness; it is what keeps the beam bright for the insulin dose. A clinician who refuses to dim it anywhere arrives at the ledge in the dark.

Match the Depth of Checking to the Potential Harm

The organizing principle is simple to state and demanding to practice: the depth of human verification should match the harm a wrong output could cause. This is risk-based verification, and it is the same logic that already governs the rest of medicine. You do not consent a blood draw the way you consent a craniotomy. You do not double-check a stool softener order the way you double-check a heparin drip. The stakes set the scrutiny, and clinicians already carry this calibration in their bones for traditional care. Risk-tiered verification simply extends that existing clinical instinct to AI outputs, which is why it should feel familiar rather than foreign.

Two dimensions determine how much harm a wrong AI output could cause, and you weigh both. The first is severity: if this output is wrong and acted upon, how badly could the patient be hurt? A wrong parking instruction wastes a trip; a wrong insulin dose can cause a seizure or death. The second is reversibility: if the error is caught later, can it be undone before harm lands, or is the consequence immediate and permanent? A misfiled administrative note can be corrected; an administered medication cannot be un-administered. High severity and low reversibility together define the outputs that demand your deepest checking, because those are the ones where a slip becomes a catastrophe you cannot take back. Low severity and high reversibility define the outputs where a lighter check is not negligence but good resource allocation.

Severity and reversibility, worked as a pair

It is tempting to collapse these two dimensions into a single vague sense of "how scary," but they are genuinely distinct, and outputs can score high on one and low on the other. A critical lab value flagged by an AI triage tool is high severity but often high reversibility: if the human misses it now, the value is still sitting in the record for the next clinician to catch, and the harm has not yet been dealt. An AI-suggested one-time IV push of a high-alert drug is high on both, because once it is given there is no next clinician and no undo. A routine progress note is low severity but, if it seeds a wrong problem onto the problem list, surprisingly low reversibility, because that error can travel silently into future decisions. Weighing both dimensions, rather than one, is what stops you from under-checking an output that looks mild but locks in permanently, and over-checking an output that looks dramatic but stays fully correctable.

Trying to check everything with equal depth is not caution; it is how you run out of caution before the dangerous output arrives. Spend your scrutiny where the harm lives.

A Tiered Verification Scheme

The practical tool is a tiered scheme: a small number of named risk levels, each with a defined depth of verification, so that the level of checking is decided by policy and habit rather than reinvented output by output under pressure. The exact tiers can be adapted to your setting, but a workable three-tier scheme looks like this.

Low stakes: a light check

These are outputs where an error is minor and easily reversible: an administrative inbox draft about parking or paperwork, a routine appointment reminder, an internal summary that no one acts on clinically without independent confirmation. A light check is appropriate: a quick scan for anything obviously wrong or inappropriate, then proceed. Spending deep scrutiny here is a misallocation, because the harm of an error is small and the cost of catching it later is trivial. Note the boundary, though: the moment an ostensibly low-stakes output starts carrying clinical content, as the inbox assistant did in the previous lesson, it is no longer low-stakes, and the tier must rise with it.

Medium stakes: a substantive check

These are outputs that influence clinical care but are not immediately dangerous and remain correctable: a routine progress note, a non-urgent care-gap suggestion, a draft that a clinician will build on before anything reaches the patient. A substantive check is warranted: read the whole output, verify the clinically meaningful claims against the source, correct what is wrong. Not the deepest possible scrutiny, but genuine engagement, because an error here can propagate into care if it is not caught, even though it is unlikely to cause immediate catastrophic harm.

High stakes: a deep check

These are the outputs where a wrong answer, acted upon, could seriously harm or kill, and where the action may be hard or impossible to reverse: a medication dose, especially a high-alert drug like insulin, anticoagulants, or opioids; a diagnosis or a diagnostic conclusion that will steer treatment; a discharge decision; anything touching a critical value or a life-sustaining intervention. Here the deepest verification is mandatory and non-negotiable: verify against every relevant source, apply full independent clinical judgment, confirm the output makes sense for this specific patient and their specific physiology, and treat the AI output as a suggestion to be earned, not a recommendation to be trusted. This is where your carefully conserved vigilance is meant to be spent, and the whole point of going light elsewhere is to have it available here.

A worked tiering table

Reading the three tiers as prose is one thing; seeing example outputs sorted by severity and reversibility is another. The table below maps concrete AI outputs to the two dimensions and the resulting check depth. Treat it as an illustration of the reasoning, not a fixed list to memorize, because the tier follows the actual content, and your unit's governance may place a given output differently.

Read the table top to bottom and the logic becomes visible: as severity climbs and reversibility falls, the required depth rises with it. The two administrative rows sit in the low tier because a mistake costs a trip or a slot, not a patient. The two middle rows earn a substantive check because an error can quietly propagate into care. The bottom three rows demand a full independent check because they combine real harm with a shrinking window to undo it. The point of writing this down is not to spare you thought at the bedside; it is to make the routine cases automatic so your judgment is fresh for the genuinely ambiguous ones.

The Gate Is Depth, Not Just Presence

It is worth being precise about what a "verification gate" actually varies as the risk changes, because the naive reading is that low-stakes outputs skip the human and high-stakes outputs get one, and that is not quite right. In a properly designed clinical workflow, a human is in the loop across the tiers wherever the output touches the patient or the record; what changes between tiers is not usually whether a human is involved but how deeply that human checks. The gate is always there for anything clinical. What the risk level sets is the depth of the check the gate performs. A low-stakes gate is a quick pass; a high-stakes gate is a full independent verification. Collapsing the two ideas, presence and depth, is a common error that leads people to think risk tiering means letting some clinical outputs through unchecked, which it does not.

This distinction matters because it keeps risk tiering from becoming an excuse. The purpose of going light on low-stakes outputs is not to remove human accountability from them; it is to spend a proportionate amount of a finite resource so that the disproportionate amount remains available where it is needed. The clinician who sends the parking reply after a quick scan is still accountable for it; they have simply, correctly, judged that a quick scan is the appropriate depth for content whose worst outcome is a wasted trip. If that same channel started carrying clinical content, the depth would have to rise, because the depth tracks the stakes of the actual output, not the convenience of the clinician. Risk tiering is a discipline for allocating verification depth wisely, not a license to stop verifying. Keep those two ideas distinct and the scheme stays honest; blur them and it becomes a rationalization for cutting corners on things that only look low-stakes.

When a gate is too loose: a dosing instruction that reached a patient

Consider how this goes wrong when a gate is tiered too loosely, told as a before-and-after. Before: a clinic runs an inbox assistant that drafts replies to patient portal messages. Because the channel is labeled "administrative messaging," the unit tiers every output from it as low-stakes, meriting only a quick scan. For months this is fine, because the messages really are about parking, forms, and hours. Then a patient messages asking what to do about a blood sugar that has been running high, and the assistant, trying to be helpful, drafts a reply that includes a specific instruction to increase the evening insulin. The clinician, conditioned by the low-stakes label on this channel, gives it the same three-second glance she gives every parking reply, does not independently reason about the dose against the patient's regimen and renal status, and sends it. The dosing instruction reaches the patient. That is a high-alert dosing decision that slipped through a gate tiered for parking.

After, with the tier attached to content rather than channel: the same message arrives, but the workflow flags that the draft now contains a dosing instruction, which pushes the output out of the low tier regardless of the channel it came through. The clinician is prompted to give it a full independent check. She stops, reviews the patient's current regimen, glucose trend, and renal function, recognizes that a blanket "increase your insulin" is unsafe advice for this patient, rewrites the reply to bring the patient in for evaluation, and documents her reasoning. The difference between the two versions is not the intelligence of the clinician or the quality of the AI. It is whether the tier was allowed to attach to the convenient category label ("administrative channel") or to the actual clinical weight of the specific output. A gate that is too loose does not announce itself; it lets harm through precisely on the day the usually harmless channel carries something that matters.

Who Sets the Tiers, and Why It Should Not Be You Alone at 2 a.m.

One of the quiet strengths of a tiered scheme is that it moves the risk judgment out of the exhausted individual moment and into a calmer, collective, prior decision. Deciding in the abstract, on a Tuesday in a governance meeting, that AI-suggested doses of high-alert medications always receive a deep check is easy and obvious. Deciding it fresh at 2 a.m. on a collapsing shift, when a deep check feels unaffordable, is where judgment fails. The whole point of setting the tiers in advance, ideally at the level of a unit or an organization rather than reinvented by each clinician under duress, is that the hardest decisions are made when they are easy and then simply followed when they are hard.

This is why risk tiering is not purely a personal skill but also an organizational one, and why the governance structures the program discusses at higher levels matter here. When a unit agrees on which categories of AI output are high-stakes and mandates the deep check for them as policy, it removes the demotion decision from the individual clinician on the worst night of their month. The clinician does not have to summon the discipline to resist tier creep in the moment, because the tier was set elsewhere and is not theirs to renegotiate. That is a kinder and safer arrangement than relying on individual willpower, and it mirrors how the rest of medicine already handles high-alert situations: independent double-checks on high-risk medications are not left to whether the nurse feels careful tonight; they are required by policy precisely so that fatigue cannot quietly waive them. Risk-tiered AI verification is the same idea, applied to a new class of outputs, and it works best when the tiers are a shared commitment rather than a lonely one.

How this connects to formal AI governance

Emerging governance frameworks point in exactly this direction. Guidance associated with accreditation bodies and health AI collaboratives has begun to call for a designated governance structure that formally evaluates AI tools for risk and bias before deployment, and that assigns responsibility for monitoring them afterward. Any specific requirement, adoption figure, or accuracy claim you encounter from such frameworks is a number to verify against the current source, not to repeat blindly, and none of it endorses a particular vendor. But the structural idea is durable and it is the one that matters for tiering: the question of which AI outputs are high-stakes should be answered by a standing body that reviews the tool's risk and bias profile, not improvised by whoever happens to be on shift. A tiering policy is where that governance judgment becomes concrete at the bedside. The committee decides, in the calm of a scheduled meeting, that AI-suggested high-alert doses always get a deep check; the clinician at 2 a.m. inherits that decision as a rule rather than re-litigating it while exhausted.

Why documenting your reasoning strengthens the record

There is a further reason the deep check matters, and it lives in the record rather than at the bedside. The standard of care around AI is evolving, and it can cut both ways: a clinician can be exposed for following a wrong AI recommendation that a reasonable review would have caught, and, as these tools become more established, potentially for ignoring an accurate one without a documented reason. Note carefully that the shape and force of that liability is itself something to verify with counsel in your jurisdiction, not a settled fact to repeat. What follows from it is practical and within your control: when you perform a high-tier check, briefly document what you found and why you agreed or disagreed. A note that says you reviewed the AI-suggested dose against the patient's renal function and adjusted it, or reviewed it and concurred for stated reasons, converts an invisible mental act into a defensible record. Documenting agreement is not busywork; it shows the check happened. Documenting disagreement and your override shows independent judgment was applied. Either way, the record reflects a clinician who verified rather than deferred, which is precisely the posture risk tiering is built to protect.

A Worked Example: The Same Tool, Three Outputs

Consider a hospitalist using an AI assistant that touches several parts of her day, and watch the tiering turn a flat, exhausting workflow into a survivable one.

At 9 a.m., the assistant drafts a reply to a patient portal message asking whether the clinic validates parking. This is low stakes: a wrong answer costs a few dollars. She scans it, sees it is reasonable, and sends it in seconds. She does not verify the parking policy against three sources, because doing so would be an absurd use of the attention she will need later. At 11 a.m., the assistant drafts a routine progress note for a stable patient. This is medium stakes: she reads it fully, confirms the assessment and plan match her reasoning, checks that no exam finding was fabricated, corrects one small inaccuracy, and signs. Genuine engagement, but not the deepest possible dive.

At 2 p.m., the assistant suggests an insulin dosing adjustment for the patient with fluctuating glucose and renal impairment. This is high stakes and low reversibility: once that insulin is given, it cannot be recalled, and an error could cause a dangerous hypoglycemic event. Here she stops and does the deep check. She verifies the current glucose trend, the renal function, the patient's intake, and the existing insulin regimen against the chart herself. She reasons independently about whether the suggested dose makes sense for this specific patient, and she notices that the AI's suggestion did not appear to account for the declining renal function that would prolong the insulin's effect. She overrides it, orders a more conservative dose, and documents her reasoning. Notice the arithmetic of vigilance: she could afford this deep, careful check at 2 p.m. precisely because she did not squander her attention on the parking reply at 9 a.m. The tiering did not just protect the insulin decision; it made protecting it possible.

Tiering an always-verify list on a real unit

Watch what it looks like when a unit builds the high tier deliberately rather than discovering it case by case. A nurse leader, a hospitalist, and a CMIO sit down with a list of every AI output their tools produce and ask one question of each: what is the worst plausible consequence if this is wrong and acted upon? They move quickly through the easy ones. The appointment reminder and the parking reply go to low. The progress note and the non-urgent care-gap suggestion go to medium. Then they reach the outputs that anchor the high tier and they name them explicitly as an always-verify list: any AI-suggested dose of insulin, an anticoagulant, or an opioid; any AI-suggested electrolyte replacement such as potassium; any sepsis or deterioration alert that would trigger escalation; any discharge-readiness flag; any handling of a flagged critical value. For each, the rule is the same, a full independent check by a clinician, and for the high-alert medications an independent double-check by a second qualified person, matching how the unit already handles those drugs outside of AI. The value of naming the list in advance is that no one has to decide, on the hard night, whether the potassium suggestion "counts." It was decided in the calm room, and the tired clinician simply follows it.

The Pitfalls of Tiering, and How to Avoid Them

Risk tiering is powerful, but it can be misapplied in ways that reintroduce the danger it was meant to remove. The first pitfall is mis-tiering: putting a genuinely high-stakes output in a low tier because it usually looks routine. Most insulin doses are unremarkable, which is exactly the trap; the tier is set by the worst plausible consequence of an error, not by how the output usually looks. If a category of output can ever cause serious, hard-to-reverse harm, it belongs in the high tier every time, not just on the days it looks scary.

The second pitfall is tier creep under pressure: on a brutal shift, the temptation is to quietly demote high-stakes checks to save time, to give the insulin dose a medium-stakes glance because the census is impossible. This is precisely when the discipline matters most and precisely when it is most likely to fail, which is why the high-tier check should be a fixed rule, not a judgment call renegotiated on every hard shift. The whole value of deciding the tiers in advance is that the decision does not depend on how you feel at 2 p.m. on a terrible day. The third pitfall is forgetting that a low-stakes output can become high-stakes the moment its content changes, which is why the tier attaches to the actual content and consequence of a specific output, not merely to its category label. A tiering scheme is a tool for spending finite vigilance wisely, and like any tool it protects you only if you apply it honestly, especially on the shifts when honesty is hardest. Done well, it is the difference between a clinician who is carefully attentive exactly where it counts and one who is uniformly, dangerously tired everywhere.

There is a fourth, quieter pitfall worth naming: letting the tool's reliability lower your tiers by stealth. A high-stakes AI that has produced good dosing suggestions for months exerts exactly the pull we studied under automation bias, and the temptation is to reason, without ever quite deciding it, that a tool this reliable no longer needs the deep check. This is tier creep produced not by a bad shift but by a good track record, and it is more insidious because it feels like earned trust rather than fatigue. The defense is the same: the tier is set by the consequence of an error, not by the tool's usual accuracy, and a high-alert dose remains a high-alert dose no matter how many times the AI has been right about it. The reliability of the tool is precisely what makes the rare wrong suggestion dangerous, because it arrives to a clinician the tool has trained to relax. Hold the high tier fixed against that pull, and you keep the deep check alive for the one suggestion in a thousand that would otherwise slip through on the strength of the nine hundred and ninety-nine that came before it.

Key Takeaways

• Vigilance is a finite budget. Checking every AI output at maximum depth is not maximum safety; it exhausts your attention on trivial outputs so it is depleted when the dangerous one arrives.
• The organizing principle is risk-based verification: the depth of human checking should match the harm a wrong output could cause. This extends the calibration clinicians already use, consenting a craniotomy differently than a blood draw.
• Weigh two dimensions of potential harm: severity (how badly could the patient be hurt) and reversibility (can the error be undone before harm lands). High severity plus low reversibility demands the deepest checking.
• Use a tiered scheme so the depth of checking is set by policy and habit, not reinvented under pressure: low stakes gets a light check, medium stakes a substantive check, high stakes a deep, non-negotiable check.
• High-stakes outputs, doses of high-alert drugs, diagnoses, discharge decisions, critical values, get full independent verification against every relevant source, treating the AI output as a suggestion to be earned, not trusted.
• Going light on low-stakes outputs is not negligence; it is what makes the deep check on the insulin dose possible. The tiering does not just protect the high-stakes decision, it makes protecting it affordable.
• Set tiers in advance through governance, and let the tier attach to an output's actual content and consequence rather than its category label, since a low-stakes channel becomes high-stakes the moment it carries a dosing instruction. Document what you checked and why you agreed or overrode.
• Guard against tier creep, whether from a brutal shift or a good track record: the high-tier check must be a fixed rule, and a high-alert dose stays high no matter how reliable the tool has seemed.