There is a spectrum of prescribing complexity in general practice, and AI is useful at different points along it.
At one end: the simple BNF lookup. A patient needs amoxicillin for a chest infection. You check the dose, confirm no allergy, and prescribe. The BNF gives you everything you need. AI adds nothing.
At the other end: the multimorbid prescribing decision. A 72-year-old with heart failure, type 2 diabetes, CKD stage 3b, gout, and depression is on eight medications. You need to add an SGLT2 inhibitor (guideline-recommended for the heart failure and diabetes), but it may affect renal function (which is already impaired), and the patient is also on a diuretic (volume depletion risk), an NSAID PRN for gout (which should probably be stopped), and an SSRI (which has a sodium interaction you need to check). Five guidelines apply. Twenty-three BNF entries are potentially relevant. And the patient is sitting in front of you.
Between these two extremes lies the territory where AI prescribing tools are genuinely useful — and where their limitations matter most.
What AI Can Automate
Guideline reconciliation. Tools like Medicaite's MetaGuideline can harmonise recommendations from multiple NICE guidelines for a single patient scenario, identifying which prescribing recommendations apply, which conflict, and which take priority. This is computational work that a clinician can do manually but that takes significant time — especially for patients with four or more comorbidities.
Interaction screening. AI can cross-reference a patient's medication list against known interactions faster than a manual BNF check. This is particularly valuable for polypharmacy patients where the interaction space is large.
Deprescribing identification. Surfacing medications that may no longer be indicated, that have been on the repeat list for years without review, or that carry accumulating risk in an ageing patient.
Guideline-linked reminders. Flagging that the patient's statin has not been escalated to the intensity recommended post-MI, or that ACE inhibitor monitoring bloods are overdue.
What AI Cannot Automate
Patient-specific contextual judgement. The guideline says to add an SGLT2 inhibitor. But this patient has a history of recurrent UTIs, lives alone, has difficulty accessing the bathroom, and expressed anxiety about starting new medications. The prescribing decision is not just pharmacological — it is personal. AI cannot weigh these factors because they are not coded in any guideline.
The conversation about medication. Shared decision-making requires explaining the options, exploring the patient's preferences, negotiating a plan they will adhere to, and safety-netting for side effects. This is a human skill that no tool can replicate.
The "I know this patient" factor. An experienced GP who has managed this patient for fifteen years carries contextual knowledge that no electronic record fully captures: previous medication intolerances that were never coded, family dynamics that affect adherence, patterns of presentation that suggest specific vulnerabilities. This knowledge shapes prescribing judgement in ways that AI cannot access.
The accountability. The prescription is the clinician's. The adverse event is the clinician's. The fitness-to-practise hearing, if it comes, is the clinician's. AI shares none of this. Professional accountability cannot be delegated to a tool, no matter how sophisticated its guideline logic.
Where iatroX Fits
iatroX sits between the BNF lookup and the prescribing copilot. It does not harmonise guidelines in the way MetaGuideline does. But it provides rapid, citation-first clarification of individual guideline recommendations — the building blocks of the prescribing decision.
When you need to check whether the NICE recommendation applies to this specific patient's eGFR, Ask iatroX gives you the answer with a link to the guideline. When you want to understand why the recommendation exists, Brainstorm walks you through the clinical reasoning. When you want to retain the prescribing logic for future patients, the Q-Bank tests your knowledge through spaced repetition.
The prescribing copilot gives you the composite answer. iatroX gives you the individual guideline components, the explanation, and the learning. The BNF gives you the definitive drug information. The clinician provides the judgement.
Conclusion
AI can automate guideline reconciliation, interaction screening, and deprescribing identification. It cannot automate contextual judgement, therapeutic relationships, or professional accountability. The clinician who understands where automation helps and where it must stop will use prescribing AI most effectively.
The right stack: MetaGuideline for complex harmonisation. iatroX for guideline clarification and learning. The BNF for definitive prescribing. And your own clinical judgement for everything that no tool can encode.