There is a spectrum of prescribing complexity in general practice, and AI is useful at different points along it.
At one end is a product-specific verification task: identify the exact medicine and check its licensed dose, contraindications or interactions in the relevant SmPC on eMC. AI may shorten retrieval, but it does not change the need to confirm the source and apply it to the patient.
At the other end is multimorbidity: several conditions, medicines, guidelines, local pathways and patient priorities interact. A tool may surface possible conflicts or organise the evidence, but it cannot reduce that decision to a pile of medicine entries. The clinician must assess benefit, harm, treatment burden, monitoring, frailty, adherence and the patient’s goals.
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. A validated system can screen a long medication list quickly, but every signal still needs clinical interpretation. Inspect the relevant products’ SmPCs, use SPS, Stockley’s or specialist advice as appropriate, and prioritise interactions in the patient’s actual context.
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 source retrieval and the prescribing decision. It provides rapid, citation-first clarification grounded in applicable guidance and medicines information from relevant SmPCs on eMC. It does not make the patient-specific decision or harmonise every competing guideline automatically.
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.
A synthesis tool can organise the relevant recommendations. iatroX can locate and explain cited source material. SmPCs on eMC provide the authorised product position. NICE, SPS, local and specialist guidance support recommendations and implementation. The clinician supplies the patient-specific 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.
A safer stack uses synthesis tools to organise complexity, iatroX for cited clarification, relevant SmPCs on eMC for product-specific information, applicable NICE, SPS, local and specialist guidance for recommendations, and professional judgement for the decisions no tool can encode.