The honest promise first: complex prescribing decisions cannot be fully verified in 90 seconds, and this article will not pretend otherwise. What can be done in 90 seconds, reliably and trainably, is a screen: a six-stage check that either clears an answer for the low-stakes use it was asked for, or identifies that a deeper review is required and names which kind. That distinction, screen versus review, is the whole method, and prescribers who install it stop making the two symmetric errors this cluster keeps meeting: treating every AI answer as terminal, and treating every AI answer as radioactive.
The six-stage screen
Question: has the AI answered the question actually asked? Drift is the quietest failure, a monitoring question answered with an initiation summary, a product question answered at class level, and fifteen seconds of comparison catches it. Jurisdiction: is this UK guidance and UK product information? The wrong-country answer is fluent by construction, and the stamp is non-negotiable for anything guidance- or licensing-shaped. Patient: are age, pregnancy, renal function, hepatic function, frailty and allergies visibly accounted for, or has the answer silently assumed the uncomplicated adult? An answer that never mentions the characteristics you supplied has not used them. Product: is the exact product, formulation and route clear, or is this an active-ingredient answer to a product-level question, the gap treated in full at /blog/product-specific-prescribing-active-ingredient-not-enough. Source: does the cited source, opened, directly support the material claim, the fidelity check from /blog/ai-citation-does-not-make-prescribing-answer-safe, run on the one load-bearing claim rather than the whole bibliography. Action: does the answer carry its consequences, monitoring, review, safety-netting, escalation, or does it stop at drug and decision, the amputation the monitoring framework exists to repair?
The red flags that end the screen and start the review
Certain territories promote any answer past screening automatically, because their error costs outrank any efficiency argument: paediatric and weight-based dosing; pregnancy and breastfeeding; renal or hepatic impairment; anticoagulants; immunosuppressants; controlled drugs; unlicensed or off-label territory; multiple interacting medicines; and any conflict with local policy. An answer touching these gets the full treatment, exact SmPC opened, guideline consulted, local document checked, and specialist or medicines-information advice where the framework's escalation logic fires, with the screen's only job being the routing. The list is worth memorising as territory, not as rules: these are the places where the gap between plausible and safe is widest, and where the 90-second promise honestly expires.
Installing the habit
Three implementation notes from how verification actually survives clinical pressure. Run the screen aloud or on paper for the first fortnight, six words, question, jurisdiction, patient, product, source, action, until the sequence is motor memory; trained users report the full pass settling near a minute for routine answers. Screen at the point of intended use, not at the point of reading, because answers consulted for orientation need only the first two stages, and the full screen is triggered by the moment an answer approaches a decision. And log the failures: every answer the screen catches, drifted, foreign, product-blurred, unsupported, goes in a one-line log, which over a month becomes your personal evidence base about which tools earn which trust for which questions, better calibrated than any published review, and, recorded with reflection, exactly the improve-practice portfolio material the competency framework rewards.
Frequently asked questions
Is 90 seconds realistic in a ten-minute consultation?
For the screen, yes, and consultations are exactly why the routing matters: a screened answer used for orientation is safe at speed, and anything the red flags catch was never a within-consultation decision anyway, which the screen makes explicit rather than discovering later.
Which stage do experienced prescribers skip most, and at what cost?
Product, because experience breeds active-ingredient thinking; the cost concentrates in formulations, devices and concentrations, which is precisely where product-level errors live.
Does the screen apply to human-written summaries too?
Identically, and running it on everything is the point: the six stages are source-agnostic verification, AI merely made the volume of summaries large enough to need a fast version.
Can the screen be delegated to the tool itself?
Partially and never fully: well-designed tools pre-answer jurisdiction, dating and sourcing by construction, which shortens the screen, and the question, patient, product and action stages are yours, because they compare the answer with a situation only you can see.
What happens when an answer passes the screen but still feels wrong?
The feeling wins: clinical unease is data the checklist cannot capture, and the correct response is promotion to full review or a colleague conversation, with the episode logged, since a screen exists to catch failures, never to overrule judgement.
Does the screen change for answers I wrote myself last month?
Only in humility: your own saved summaries age exactly like anyone else's, and the jurisdiction and source stages catch the update you have not yet heard about, which is the recency problem applied to your own notes.