The question splits into seven tasks the moment it is taken seriously, and AI's reliability differs across them, which is why blanket answers in either direction mislead. But one structural fact governs the whole territory and deserves the opening sentence: in interaction checking, omission is the dangerous failure, and an omission is precisely what the user cannot verify, because you cannot inspect a warning the system never raised. That asymmetry, checkable errors versus uncheckable silences, is what makes AI a legitimate organiser of medication reviews and an illegitimate sole screen, and everything practical follows from it.
The seven tasks, separated
Medication transcription: converting lists between formats and records, where AI is capable and where every error it introduces propagates downstream, so transcribed lists get read back against the source. Deduplication: catching the same medicine present twice, including the brand-generic pair that fools string matching, a task AI performs usefully with product-level knowledge. Indication mapping: pairing each medicine with its reason, which surfaces the indication-less legacy items and is among AI's most genuinely helpful review functions. Interaction identification: the flagged-pairs task, where AI output is a prompt list for verification against established interaction resources and pharmacist review, never the screen itself, per the omission asymmetry, and where a proof-of-concept literature exploring LLMs for medication review exists alongside the standing finding that specialist human review remains necessary. Contraindication screening: patient-factor checks that are only as good as the patient data supplied, which is usually the binding constraint. Deprescribing suggestions: hypotheses for clinical judgement, the polypharmacy article's territory, /blog/ai-multimorbidity-polypharmacy-anp. And reconciliation proper: resolving the differences between what the record says, what was dispensed, what the discharge letter changed and what the patient actually takes, which is not a data task at all but an inquiry, and the task the word "reconciliation" actually means.
The data-quality failures that corrupt everything upstream
Any AI pass inherits its inputs, and medication data fails in patterned ways worth checking deliberately: brand and generic duplicates split across list sections; the wrong formulation recorded, modified-release lost in transcription; PRN medicines omitted from "regular" lists and doing interactions anyway; over-the-counter and herbal products absent from every record and present in the patient; hospital-only and community-only records each holding half the truth; and the patient taking a different dose, frequency or nothing at all relative to the record, the gap only conversation finds. The operational rule: an AI-organised list is reconciled against at least the patient's own account and the repeat list, plus discharge information and dispensing history where available, before any conclusion drawn from it is trusted, because analysis of an incomplete list is confident analysis of a fictional patient.
The stress test, and the honest verdict
Run any reconciliation tool, or your own workflow, against the classic case: a patient on methotrexate whose acute presentation tempts a trimethoprim prescription, the interaction that recurs in serious-incident literature precisely because it hides across settings and list fragments. The test is not whether a system knows the interaction, most do, but whether the workflow guarantees the two medicines ever meet in the same checked list: the methotrexate on the repeat list, the trimethoprim in the acute moment, the PRN and hospital fragments reconciled, the patient asked. That is the honest verdict in one case: the failure mode is architectural, fragmented truth, more than informational, and AI helps exactly insofar as it assembles fragments for a clinician who then verifies against established interaction resources and, for complex regimens, pharmacist review, the second-look configuration this platform claims for itself and no more: structure and source navigation, never the sole interaction database and never the record.
Frequently asked questions
Which established resources should the verification use?
The SmPC's interaction sections for the products in question, recognised interaction references, and the pharmacist, with agreement across layers for consequential combinations; a single silent layer is never clearance, per the omission asymmetry.
Is AI better at reconciliation in hospitals or primary care?
The data differs more than the AI does: each setting holds different fragments, and the tool's value tracks how much of the truth its inputs actually contained, which is why the patient conversation outranks every integration.
What should be documented after an AI-assisted review?
The sources of the reconciled list, the checks run, the discrepancies found and resolved, and the changes with rationale and review dates, ordinary medication-review documentation, with the AI's organising role noted honestly.
Can patients help close the reconciliation gap directly?
They are the gap's other half: the brown-bag review, everything actually taken brought and walked through, remains the single highest-yield reconciliation act, and AI's list organisation makes the conversation faster, never unnecessary.
Should reconciliation AI be a separate tool or part of the record system?
Wherever it lives, the tests are identical: what fragments it sees, how it handles the products versus ingredients problem, and whether its output arrives labelled as a checked screen or an organising draft, because the label decides how safely it will be used.