Dear Abby PharmD, Should I second-guess AI's Every Move?

Dear Somewhere Between,

That catch you made last week, that wasn't a coincidence. That was clinical expertise doing exactly what it's supposed to do — and it's a perfect illustration of why the right relationship with AI isn't blind faith or reflexive suspicion. It's calibrated trust.

Here's the problem with becoming the AI's yes-man: clinical decision support tools, however sophisticated, are trained on population-level data and structured inputs. They excel at pattern recognition across thousands of variables simultaneously — but they don't know your patient. They don't know she stopped taking the interacting medication three weeks ago, or that his renal function has been quietly declining despite a "normal" last creatinine. You do. That gap between the algorithm's dataset and the individual in front of you is precisely where pharmacist judgment lives.¹

This isn't a hypothetical risk. Studies on alert fatigue — a well-documented phenomenon in both pharmacy and physician workflows — show that high alert volumes lead clinicians to override or dismiss warnings without adequate evaluation, and that the same habitual clicking can extend in reverse: uncritical acceptance of AI recommendations that seem right but lack the full clinical picture.² One analysis found that clinicians accepted AI-generated recommendations at significantly higher rates when the recommendation was presented with confident, authoritative language — regardless of whether it was actually correct.³ The interface design itself can nudge you toward over-trust.

So what does calibrated trust look like in practice? Think of it as "trust but verify" with some structure behind it. Accept that the AI is genuinely excellent at certain tasks — high-volume, rule-based screening; renal and hepatic dose flagging; formulary and cost checks. For these, reasonable deference makes sense and preserves your cognitive bandwidth for harder problems. But for complex patients, polypharmacy cases, or any recommendation that would represent a significant clinical change, treat the AI output as a starting point, not a conclusion. Ask: “What is the reasoning? Does it account for this patient's full context? Would I reach the same place without the prompt?”⁴

Professional pharmacy organizations have begun addressing this directly. Emerging competency frameworks for AI in practice emphasize that pharmacists must maintain independent clinical reasoning skills precisely because automation can erode them over time — a concept sometimes called "automation bias."⁵ The concern isn't that the tool is wrong often; it's that the times it iswrong may be the highest-stakes moments, and an undertrained override reflex is a dangerous thing to bring to those situations.

Your technician's joke stings a little because it contains a real warning. The pharmacists best positioned to use AI well are those who can articulate why they agree with a recommendation, not just that they do. Stay curious about the rationale. Keep the habit of occasionally working through clinical problems independently, before you see what the algorithm says. And when your gut says something doesn't fit — as it did last week — follow it.

The goal isn't to be faster than the AI. It's to be wiser than it. That's a job it still can't do.

— Abby PharmD