An oncology pharmacist named Priya faced a decision late one afternoon that no prompt would have improved. A patient on a narrow-therapeutic-index chemotherapy agent had a borderline lab result, a complex history, and a question about whether to hold the next dose, where one wrong call in either direction carried real risk: hold unnecessarily and the cancer advances, dose into a dangerous level and the patient is harmed. By 2026 Priya used AI fluently for prior authorization (PA), counseling drafts, and verification support, and her instinct was to ask the assistant to weigh the factors. She stopped herself. This was not a task where a fast, plausible draft helped; it was a high-stakes clinical judgment with no margin for a confident-but-wrong answer, where the very thing AI is good at, producing a fluent, reasonable-sounding recommendation, was the thing most likely to mislead her under time pressure. She made the call herself, consulted the prescriber, and documented her reasoning. The point of this lesson is not that Priya distrusts AI; she uses it constantly. The point is that a mature AI practitioner knows, in advance and in writing, the specific decisions where AI should not be used at all, because the failure mode is catastrophic and the tool's fluency is a liability rather than a help. This is the kill-criteria list: the line, drawn before the pressure arrives, that says here the tool stops.

Why You Need a Line Drawn in Advance

The most important word in this lesson is "in advance." The danger is not that a pharmacist would consciously decide to hand a life-or-death judgment to a chatbot; almost no one does that on purpose. The danger is drift: a tool you use successfully a hundred times a day for low-risk tasks becomes the reflex you reach for on the hundred-and-first, which happens to be high-risk, and the reach is automatic before you have consciously weighed whether it belongs there. Under time pressure, with a queue backing up and a fluent tool one keystroke away, the decision to use AI is rarely a decision at all; it is a habit. A written kill-criteria list interrupts that habit by deciding the boundary once, calmly, before any specific case, so that in the moment you are not relying on your judgment-under-pressure to notice you have crossed a line you never explicitly drew.

This is why the boundary must be written, not merely felt. A felt boundary moves with fatigue, confidence, and deadline pressure; it is most likely to give way exactly when the stakes are highest, because that is when you are most rushed and most tempted by a fast answer. A written boundary does not move. It says, in advance and in plain language, "for this category of decision, the tool is not used, regardless of how convenient it would be in the moment," and it lets a tired pharmacist at the end of a shift defer to a rule made by a clear-headed one. The kill-criteria list is the pharmacy's pre-commitment, the same logic as any safety stop: you decide where the hard limit is when you are calm, so the limit holds when you are not.

A kill-criteria list is the set of decisions, written in advance, where AI is not used at all. Its purpose is to stop drift: to draw the boundary calmly before the pressure arrives, so a fluent tool does not become the reflex on a decision where being plausibly wrong is catastrophic.

The Asymmetry That Defines the Line

To know where to draw the line, you have to understand what makes a decision wrong for AI, and it comes down to an asymmetry this program has built on from the start. For most tasks, the cost of an AI error is bounded and catchable: a wrong PA criterion gets caught in verification, a softened warning gets fixed before the patient hears it, a bad inventory forecast gets corrected. The tool drafts, you verify, and the verification catches the error before it reaches a patient. The kill criteria apply to a different class of decision, where that safety net is absent or unreliable: decisions where a plausible-but-wrong answer can act before verification catches it, where verification itself is the hard clinical judgment rather than a separate check, or where the consequence of a single wrong call is severe and irreversible. In those decisions, the model's greatest strength, generating a fluent, confident, reasonable-sounding answer, becomes its greatest hazard, because it produces something that feels checked when the checking is the very thing only a clinician can do.

Priya's case sits exactly there. The decision to hold or give that chemotherapy dose was not a draft she could verify against a source; it was the clinical judgment itself, weighing a borderline lab, a complex history, and an irreversible downside in both directions. An AI recommendation would not have been a draft to check, it would have been an answer that felt authoritative and could anchor her judgment toward a plausible wrong call, with no separate verification step to catch it because the judgment was the verification. That is the signature of a kill-criteria decision: the AI cannot help because there is no safe division of labor between a draft and a check. The whole task is the check, and the check is irreducibly human.

Contrast this with where AI genuinely belongs, because the line is not "AI is risky, avoid it." Used on the right tasks, with verification, AI is a major patient-safety and access win; the program exists because collapsing PA turnaround gets patients on therapy faster. The kill-criteria list does not retreat from that. It sharpens it, by carving out the specific decisions where the draft-then-verify model breaks down, so that the tool stays powerful where it helps and absent where it would do harm. A mature practitioner is not the one who uses AI everywhere or nowhere; it is the one who knows precisely which decisions are on which side of the line.

What Belongs on the Kill-Criteria List

A useful kill-criteria list is concrete, naming categories a pharmacist can recognize in the moment. The exact list belongs to each pharmacy and its scope, but several categories recur, and they share the asymmetry described above. High-risk, narrow-margin clinical judgments belong on it: decisions about narrow-therapeutic-index drugs, chemotherapy dosing, and similar calls where a plausible wrong answer is severe and the judgment cannot be reduced to a verifiable draft. Irreversible or emergency decisions belong on it: situations where there is no time to verify and no way to undo a wrong call, so a fast plausible answer is precisely the trap. Decisions that are themselves the verification belong on it: where the clinical judgment is the safety check, so there is no separate step in which an error would be caught.

The list extends beyond the purely clinical. Final accountability decisions belong on it: the clinical sign-off, the assertion that goes to a payer or into the record under the pharmacist's license, is the pharmacist's to make and own, never delegated to a tool, because "the AI surfaced it" is never the clinical decision. Anything that would put unprotected protected health information (PHI) into an ungoverned tool belongs on it: a privacy line that is categorical, not a judgment call in the moment. Decisions outside the tool's grounded scope belong on it: asking an ungrounded general model a question it has no authoritative source for is asking for a confident guess, and on a clinical decision a confident guess is a kill-criteria situation. Each category names a place where the draft-then-verify safety net fails, which is what unites them and what a pharmacist should learn to recognize.

It is worth stating what the list is not. It is not a ban on AI near hard problems; a pharmacist working a complex case can still use AI to pull background, draft a counseling explanation, or assemble a PA, as long as the load-bearing clinical judgment, the kill-criteria part, stays out of the tool and in the clinician. The line runs through the task, separating the supportable drafting from the irreducible judgment, not around the whole case. The skill is to keep using AI for the parts that are safe to draft and verify, while holding the specific decision that meets a kill criterion entirely outside it. A good list helps a pharmacist make that separation cleanly, in the moment, under pressure.

A short worked example shows how the line cuts through a single case rather than around it. Suppose a patient on warfarin presents with a new prescription and a borderline lab, and the pharmacist must decide whether the combination is safe to dispense and what to counsel. Several parts of this case are squarely on the safe side of the line: AI can pull the patient's relevant history from the record, draft a plain-language counseling explanation of the new drug, and assemble the documentation, each of which the pharmacist then verifies. One part is squarely on the kill side: the clinical judgment about whether the borderline lab and the interaction make dispensing unsafe for this specific patient is a narrow-margin call where a plausible wrong answer is severe and the judgment is itself the verification. The mature practitioner uses AI for the first set of tasks and keeps the dispensing-safety judgment entirely human, consulting the prescriber and documenting the reasoning. Same case, both sides of the line, and the kill-criteria list is what lets the pharmacist tell, quickly, which part is which.

Writing and Using the List

A kill-criteria list only works if it is usable in the moment, which means it must be specific, short, and known to the team. Writing it is a deliberate exercise: gather the pharmacists, walk through the real decisions the pharmacy makes, and sort them into "draft-then-verify is safe here" and "this is a kill-criteria decision." For each entry on the kill side, state the category in language a rushed pharmacist will recognize, and state plainly what happens instead: the human makes the call, the prescriber is consulted, the decision is documented. The output is not a philosophical document; it is an operational rule that a technician or a covering pharmacist can apply without re-deriving the reasoning, because the reasoning was done once, in advance, by the team.

Using the list well takes a few habits. Make it visible. A kill-criteria list filed and forgotten protects no one; it has to be where the work happens and part of how new staff are trained, so the boundary is shared, not private. Treat a kill criterion as a hard stop, not a strong suggestion. The entire value is that it does not bend under deadline pressure, so when a case meets a criterion, the tool is out, full stop, and the pharmacist proceeds by judgment. Document the human decision. When a case meets a kill criterion, the record should show that the decision was made by the pharmacist and how, which serves the patient, the pharmacy, and the accreditation expectation that AI use, and non-use, is governed and traceable. Revisit the list as tools and scope change. A capability that was unsafe to draft last year may become a supportable draft-then-verify task with a properly grounded tool, and a list that never updates either over-restricts or, worse, fails to add new high-risk uses as they appear.

A few common objections are worth answering directly, because they are the arguments that erode a kill-criteria list in practice. The first is "but the AI is usually right, so why exclude it from the hard cases?" The answer is the asymmetry: on a kill-criteria decision the cost of the occasional wrong answer is catastrophic and uncatchable, so "usually right" is precisely not good enough, in the way an aircraft checklist is not waived because the engine usually starts. The second objection is "excluding AI just slows us down on the cases that need speed." The answer is that the kill-criteria categories are exactly the cases where speed at the cost of a wrong call is the danger, not the goal; the list does not slow down the PA queue, it protects the narrow set of irreversible judgments where a fast wrong answer is worse than a slower right one. The third is "a good system prompt and grounding make even the hard cases safe." The answer, from the earlier lessons, is that grounding and system prompts reduce error but never eliminate it, which is enough for draft-then-verify tasks and not enough for decisions where the judgment is the verification and a single error is severe. None of these objections survives contact with the asymmetry, which is why naming them in advance, when the list is written, keeps the list standing when they resurface under pressure.

That last habit deserves emphasis, because the list is a living control, not a monument. The goal is not maximum restriction; it is correct restriction, which means the line should move as evidence and tooling justify, in both directions. As a grounded PA tool proves it can safely draft a category that once felt too risky, that category can move to the verify-able side. As a new high-stakes use appears, perhaps a new class of clinical recommendation a vendor is eager to automate, it should be evaluated against the asymmetry and added to the kill side if a plausible wrong answer would act before verification could catch it. A pharmacy that revisits its kill-criteria list is doing governance: deciding, deliberately and on the record, where the tool helps and where it must stop.

The Judgment Behind the List

Underneath the operational list sits a single piece of professional judgment that is worth making explicit, because it is the thing the whole program has been building toward. The kill-criteria list is the clearest expression of the cardinal rule: AI supports the pharmacist's judgment and never replaces it. Every advanced technique in this chapter, chain-of-thought to make reasoning visible, persona engineering to shape framing, sharpens how AI supports judgment. The kill-criteria list draws the matching boundary on the other side: it names the decisions where there is no supporting role to play, where the judgment is the task and the tool's contribution would be only a more confident way to be wrong. Knowing where that boundary lies is not a limitation on AI skill; it is the highest form of it.

This is why the most advanced AI practitioner in a pharmacy is often the one most willing to not use AI on a given decision. Fluency with the tool includes fluency with its absence: recognizing, fast and under pressure, that this particular call is one the tool must stay out of, and having the discipline and the written backing to keep it out. Priya was not less skilled with AI for declining to use it on the chemotherapy decision; she was demonstrating exactly the judgment that makes AI safe to use everywhere else. The pharmacist who can both wield the tool expertly and set it down deliberately, on the decisions that meet the written criteria, is the practitioner the cardinal rule, and the coming accreditation, are describing. The kill-criteria list is how a pharmacy makes that judgment a shared, durable, governed standard rather than a private instinct that holds only until the queue gets long.

Key Takeaways

• A kill-criteria list is a written, in-advance set of decisions where AI is not used at all; its purpose is to stop drift, the tendency for a fluent tool used safely a hundred times to become the reflex on the high-risk hundred-and-first.
• The boundary must be written, not felt, because a felt boundary moves with fatigue and deadline pressure and gives way exactly when the stakes are highest; a written rule lets a calm decision govern a rushed moment.
• The line is defined by an asymmetry: kill-criteria decisions are those where a plausible-but-wrong answer can act before verification catches it, where the judgment is itself the verification, or where a single wrong call is severe and irreversible.
• On those decisions the model's strength, fluent and confident output, becomes the hazard, because there is no safe division of labor between a draft and a check when the whole task is the check.
• Categories that recur on the list: high-risk narrow-margin clinical judgments, irreversible or emergency decisions, decisions that are themselves the verification, final accountability and sign-off, putting unprotected PHI into an ungoverned tool, and questions outside a tool's grounded scope.
• The line runs through the task, not around the case: a pharmacist can still use AI to draft and verify the supportable parts of a complex case while keeping the specific kill-criteria judgment entirely human.
• Make the list usable: specific, short, visible, treated as a hard stop, with the human decision documented, and revisited as tools and scope change so it stays correctly, not maximally, restrictive.
• The kill-criteria list is the clearest expression of the cardinal rule that AI supports judgment and never replaces it; the most advanced practitioner is often the one most willing to deliberately not use AI on a given decision.