The patient this article is about does not appear in any single guideline: diabetes, chronic kidney disease, heart failure, osteoarthritis, depression and a history of falls, each condition trailing its own evidence-based recommendations, which in combination produce conflicts, burdens and interactions no individual guideline acknowledges, because each was written as if its condition arrived alone. This is the terrain where ANP medication review lives, and the honest question about AI is narrow and useful: can it help organise the collisions? Organise, yes, plausibly well; resolve, no, and the difference is the whole article, because a 2025 qualitative study of English primary-care clinicians exploring AI-supported prescribing for multiple long-term conditions found exactly this pairing, perceived value alongside concerns about integration, professional responsibility, guideline conflict and effects on the clinical relationship.
The seven sorting categories
The productive use of AI in a complex review is as a structuring instrument: given the medication list and conditions, sort the territory into categories a clinician then walks. Immediate safety problems: combinations or continuations posing current harm risk, the category that jumps the queue. Medicines without a current indication: the legacy items whose original reason has resolved, moved or was never recorded. High-risk interactions: flagged for verification against established interaction resources, never accepted from any single system, since the omission an AI fails to mention is unverifiable by definition. Monitoring gaps: medicines continuing without the surveillance their use assumes. Therapeutic duplication: same effect, multiple agents, including the brand-generic duplications reconciliation misses. Treatment burden: the aggregate load of doses, timings and appointments the regimen imposes on a person. And patient priorities: what this patient wants treatment to achieve, which no data field holds and every decision should serve. An AI pass that populates these seven has genuinely helped, not by deciding anything, but by converting an overwhelming list into a walkable agenda.
Why the optimised list is never implemented wholesale
The failure mode worth naming precisely: an AI-generated "optimised regimen" is a set of hypotheses wearing a plan's formatting, and implementing it wholesale imports every limitation at once, guideline targets treated as independently mandatory when frailty, life expectancy and preference argue otherwise; interactions assessed without the over-the-counter and herbal products the record never held; stopping suggestions blind to withdrawal, rebound and the difference between a medicine that is unnecessary and one that is unsafe to remove abruptly. The clinician's frame corrects each: frailty and functional status modulate every target; time-to-benefit gets weighed against realistic horizons; the patient's own report reconciles the list the record believes; and every change carries a review plan. The distinction the qualitative literature's clinicians drew is the right one, value in the organising, danger in the deciding, and the responsibility question resolves the same way it always does: the accountable prescriber remains the author of every change, with the AI's contribution documented as structure, not authority.
The shared-decision review template
A practical close for the fourteen-medicine consultation: safety category actioned first, with the patient told why; one or two changes per review as the default tempo, because simultaneous changes make attribution of benefit and harm impossible; each change recorded with rationale, expected effect, review date and the symptom that would prompt earlier contact; priorities revisited in the patient's words, what matters to you about this list; and the whole review captured as the CPD it genuinely is, complex medication reviews being among the richest reflective material an ANP's revalidation can draw on. Used this way, the AI never met the patient and never needed to: it organised the evidence so the consultation could be about the person, which is the only version of "seeing the whole patient" a tool can honestly claim.
Frequently asked questions
Which sorting category does AI perform worst at?
Interactions, for the structural reason above: omissions are invisible, so AI output functions as a prompt list for verification against established interaction resources and pharmacist review, never as the screen itself.
How should conflicting guideline targets be documented?
As a named trade-off with the patient's priority attached: recording that a target was deliberately relaxed for frailty or preference converts an audit anomaly into evidenced person-centred care.
Is deprescribing just this article run in reverse?
It is its own discipline with its own hazards, tapering, withdrawal, goals-of-care framing, and the frailty and care-home version of this territory deserves the dedicated treatment the wider cluster gives it.
How long should a fourteen-medicine review take with this method?
Longer than a routine appointment and shorter than it fears: the AI sorting pass compresses preparation substantially, and the consultation itself spends its time where the template puts it, priorities and one or two well-chosen changes.