Antimicrobial prescribing is the sharpest test of every principle in this cluster, because the right answer is decided by variables no global model reliably holds: local resistance patterns, care setting, precise syndrome definition, the real content of a reported allergy, renal function, and the stewardship policy your organisation has committed to. A globally plausible antibiotic answer can be locally wrong on agent, route, duration or all three, and the evidence says this is not hypothetical: a 2025 study found substantial variation among general-purpose models in antibiotic selection, dosing and duration for common scenarios, exactly the parameters where wrongness does harm.
Where antimicrobial answers go wrong
Seven failure points, worth knowing as a checklist because each is independently sufficient. Agent: the nationally reasonable choice that local resistance data has retired. Route: oral-versus-intravenous decisions that depend on severity assessment and setting, not on the syndrome's name. Duration: the parameter with the widest model variation and the strongest stewardship stakes, where shorter evidence-based courses are quietly overwritten by older habits in training data. Allergy interpretation: "penicillin allergy" spans childhood rash to anaphylaxis, and the answer that treats the label as binary either denies first-line treatment unnecessarily or risks the reaction; the recorded reaction's nature is the actual variable. Source control: the antibiotic question that is really a drainage, removal or referral question, which a medicine-shaped answer format structurally under-asks. Review and de-escalation: the stop-and-narrow discipline stewardship runs on, absent from answers that end at initiation. And the local layer itself: the trust's antimicrobial policy, which encodes resistance surveillance and formulary decisions no national or global source carries, and which operationally governs.
Reading the evidence honestly, including the encouraging part
The instructive contrast: against the general-purpose variation above, a 2026 single-site simulation study of Ask Eolas, grounded in one organisation's own approved antimicrobial guidance, found no prescribing errors in the Ask Eolas group compared with six and eight errors in the two comparator groups, across 45 healthcare professionals. Read both halves with discipline. The encouraging half: local grounding works, an answer layer constrained to the organisation's own current policy removed the failure modes that come from answering out of global distribution, which is architectural confirmation of this article's thesis. The bounded half: a controlled, single-site simulation with a small sample is evidence of promise, not of universal real-world safety, and the study's own authors call for implementation research; the honest generalisation is that grounding in current local policy is the right design, and that live deployment evidence remains the category's next requirement. For a prescriber choosing tools, the translation is direct: for antimicrobial questions specifically, the value hierarchy runs local policy engine, then national guidance, then everything else, and any tool's answer that has not seen your trust's policy is orientation, not instruction.
The local-first workflow, tested on a hard case
The worked pattern, on the case built to stress it: suspected pyelonephritis, impaired renal function, reported penicillin allergy. Syndrome and severity first, because upper-versus-lower and systemic features decide setting and route before any agent question opens. Local antimicrobial policy next, the operational answer for agent and duration in your organisation, with its allergy pathway applied to the reaction's actual recorded nature rather than the label. National guidance for the frame around it, and for anything the local document does not cover. Renal function into every step, agent suitability and any adjustment verified against the exact product's SmPC, never computed from a synthesis's memory, with the deeper renal-dosing hazards flagged for their own treatment. Escalation criteria named, and review-and-de-escalation booked at initiation, because stewardship is a follow-up behaviour, not a prescribing moment. Where does an evidence layer like Ask-iatroX fit? Exactly where this cluster always puts it: retrieving the national evidence, explaining the reasoning, surfacing the SmPC, and directing the final operational decision to current local guidance, which is not a limitation of the tool but the correct shape of the whole system.
Frequently asked questions
Why does duration vary so much across AI answers?
Training data spans decades of shifting practice, and duration recommendations have shortened as evidence accumulated; models average history, stewardship follows current guidance, and the divergence is structural, which is why duration always gets the local-policy check.
How should a reported allergy be handled in prompts?
By nature, not label: state the recorded reaction and its severity, and treat any answer that did not use that information as having answered a different patient.
Does the Eolas result mean every organisation should deploy a local answer engine?
It means locally grounded architecture has the right shape and early evidence; procurement still runs through the standard questions, currency of the uploaded corpus, behaviour when sources conflict or are absent, and live monitoring, the same rubric this series applies everywhere.