Discharge Medicines Reconciliation: The Plumbing Matters More Than the AI

The problem

The anchor paper in AI in Practice covered ambient AI scribes: tools that listen to a consultation and draft a note for the clinician to check and sign¹. The largest documented deployment recorded 7,260 physicians using them across 2,576,627 encounters between October 2023 and December 2024 (measured usage data)². Accuracy had been tested only in simulated encounters, where five platforms omitted or got wrong an average of 26.3% of key clinical elements in unedited notes (simulation, tested December 2024 to February 2025)³. This paper looks at a different problem: not drafting one clinician's note, but keeping a patient's medicines list correct as it crosses from hospital to GP and pharmacy, where the output has to be exact and other organisations act on it.

A modelling study funded by NHS England estimated that around 1.8 million medication errors a year occur at hospital transitions in England and would go undetected by standard medicines reconciliation (95% credibility interval 1.3 to 2.6 million; modelled from published literature and expert input, 2024)⁴. Of these, 44% arose at discharge and 52% at admission⁴. The authors estimated that the errors harm around 31,500 patients, use about 36,500 extra bed days costing around £17.8 million at 2020–21 prices, and contribute to more than 40 deaths, with wide uncertainty around the deaths (modelled estimate). They describe all the figures as indicative rather than definitive⁴.

Studies that followed patients after they left hospital point the same way. A systematic review of 54 studies covering 20,895 discharges in 26 countries found a median rate of unintentional medication discrepancies of 50% in adults (interquartile range 39% to 76%, 11 studies) and a median adverse drug event rate of 19% (7 studies). The studies it covered ran to March 2019, so the figures are older than three years, and the studies were too varied to pool⁵. In plain terms, in the typical study about half of the adults leaving hospital had an unintended difference or error in their medicines, and about one in five had an adverse drug event⁵.

The World Health Organization named transitions of care as one of three priority areas when it launched its Medication Without Harm challenge in March 2017, with a goal of halving severe, avoidable medication-related harm over five years. It noted that most harm arises from failures in how care is organised and coordinated, especially when several providers are involved (statement from 2017, older than three years)⁶.

What already works

Reconciliation by people. A meta-analysis of 24 randomised trials with 17,664 participants aged 65 or over found that interventions including medicines reconciliation were associated with fewer hospital readmissions (relative risk 0.88, 95% confidence interval 0.81 to 0.96). Interventions that bridged the move from hospital to home for up to 90 days worked best. The authors named subjective coding of interventions and differences between studies as limits (controlled trials, published 2020, so older than three years)⁷.

England's Discharge Medicines Service. Since February 2021, hospitals in England can refer a discharged patient to their community pharmacy, which reviews the referral within 72 hours, checks that the first post-discharge prescription matches the discharge plan, and counsels the patient⁸. At one trust, Royal Cornwall Hospitals, pharmacies rejected 3,283 of 40,259 referrals between February 2021 and December 2025 (8%, falling to 5% in 2025), and a further 7,051 (18%) were recorded as having had no action (operator-reported service evaluation, single trust). The authors estimated that up to 82% of rejections were potentially avoidable, and they know of no formal evaluation of the service's effect on patient benefit⁸. A 2022 freedom of information survey found that 39 of 147 hospital trusts (27%) had not yet started referring, and that trusts offered referrals to very different groups of patients (trust self-reports, July 2022, older than three years)⁹. In plain terms, the route that could receive a reconciled list already exists, but it is used unevenly and has not been formally evaluated for patient benefit⁸⁹.

Shared records. An information standard introduced in England in 2023 (ISN DAPB4013) aims to make medication and prescription information computable and shareable across settings. When the modelling study was written, moving the information still needed a healthcare professional to carry it across, and it was unclear how interoperable hospital prescribing systems were⁴.

Discharge summaries drafted by AI. In a blinded review at one US hospital of 100 inpatient stays from 2019 to 2022, narratives drafted by a large language model were of comparable quality to physician-written ones but contained more errors per summary (2.91 against 1.82), with low potential for harm in both (0.84 against 0.36 on a scale of 0 to 7) (search extract)¹⁰. A single-arm pilot on one inpatient unit between August and October 2025 generated 1,274 hospital-course summaries across 384 discharges, and its authors reported that the drafts were often used and that little risk of harm was identified (search extract)¹¹. In plain terms, drafting from the record is being tried in live care, but these studies concern narrative summaries rather than medicines lists, and the pilot had no comparison group¹⁰¹¹.

Medicines reconciliation by a model. A retrospective study at one Chinese cancer hospital, covering 249 patients who were taking long-term medicines before surgery (January 2024 to January 2026), tested a locally hosted large language model working from a structured knowledge base. Time spent on reconciliation fell by more than 90% compared with manual work at three levels of difficulty, and the model's recommendations scored 0.80 on a text-similarity measure (retrospective, single centre)¹². In plain terms, a model could draft a recommendation far faster, but the abstract reports text-similarity and rater scores, not errors caught or patient outcomes, and the setting was admission for surgery, not discharge home¹².

A test on synthetic patients. A preprint compared five open-weight models on 200 synthetic patient records (4,000 runs). The largest model listed active medicines almost perfectly in its best format (F1 score 0.9956), while a smaller model's recall fell to 0.24 for patients with 11 active medicines. In all 20 combinations of model and format, precision exceeded recall, so models more often missed a current medicine than invented one (synthetic data, exact name matching, preprint, date not shown)¹³. In plain terms, the models tended to leave medicines out rather than make them up, and they struggled most with the patients on the longest lists¹³.

What does not yet exist. No study was found in which an AI system compared a patient's pre-admission medicines, the changes made in hospital and the discharge prescription, and then passed a checked list to the GP and pharmacy, whether in a trial or in routine use. The evidence on the problem and on the pharmacy route comes mostly from England; the evidence on AI drafting comes from the United States, China and synthetic data.

The possibility

One way this could work is as a joined-up chain rather than a single product. During the stay, a system would keep a running comparison of three lists: the best available list of what the patient took before admission, the changes made in hospital with their reasons, and the discharge prescription. Before discharge it would flag mismatches, such as a long-term medicine that is on the home list but missing from the discharge prescription with no recorded reason. A pharmacist or doctor would resolve each flag. The confirmed list would then pass in structured form to the patient's GP record and nominated pharmacy, which could run the checks the Discharge Medicines Service already describes, and the patient would receive a plain-language explanation to check.

What depends on AI. The AI part would be narrow: reading free text such as ward notes and letters, pulling out medicine names, doses and stop or start decisions, matching them to standard codes, and drafting the patient explanation. Everything else would be ordinary systems and policy: the electronic prescribing record, the shared standard that lets systems exchange medicines data, the referral route to pharmacies, and decisions about who is paid for the checking and who is accountable for the list. If most discrepancies arise because information is re-typed or never sent, shared records could remove much of the problem before any AI is added. If instead much of the relevant information sits only in free text or conversation, AI reading could matter more. The sources found do not say how discrepancies divide between the two.

If the discharge conversation were also captured by an ambient scribe, it could supply a record of what the patient was told about changed medicines, which a later check could compare with what the GP and pharmacy received. That would take the anchor's capability from drafting a note to feeding a list that other people act on.

What would have to be true

1. The starting list would have to be right. In the modelling study, 52% of transition errors arose at admission, so a flawless comparison at discharge could still pass on a wrong list⁴.
2. Medicines data would have to move between hospital, GP and pharmacy in structured form. The modelling study explored cuts in transition errors of 10% to 50% from an interoperable record, but its effect-size evidence came from two single-centre before-and-after studies (relative risks 0.53 and 0.69) and a pooled figure of 0.55 for electronic reconciliation tools that were not all interoperable (modelled scenarios, 2024)⁴.
3. The AI would have to hold up on complex patients and on real records. Smaller models in the synthetic benchmark lost recall beyond 7 to 10 active medicines, and the author expected the figures to be upper bounds for real hospital data (synthetic data, preprint)¹³. In the narrative study, AI drafts carried more errors per summary than physician drafts (2.91 against 1.82) (search extract)¹⁰.
4. A human check would have to catch what the AI misses. Because omission was the main failure in the benchmark, reviewers would need to look for what is absent, not only what is wrong¹³. The authors of the scribe accuracy study noted that no evidence yet shows whether clinician review reliably catches such errors in practice³.
5. The tool's legal status would have to be settled. The MHRA's guidance of 29 July 2026 treats an ambient product that drafts a discharge summary, or formats medication lists for clinician review and confirmation, as not a medical device when it gives no new clinical information, and treats a product that analyses information and acts without clinician confirmation as a device¹⁴. The guidance does not mention medicines reconciliation. If a reconciliation tool recommended which medicines to continue or stop, it could fall on the device side, and the guidance does not say¹⁴.
6. The receiving end would have to act on the list. At the Cornwall trust, 18% of referrals were recorded with no action, and 57% of the potentially avoidable rejections were process failures at the community pharmacy, such as not actioning referrals in time; the authors point to staffing and capacity strain in community pharmacy (operator-reported service evaluation, to December 2025)⁸.

What could stop it

What the evidence shows

An evidence gap. No study was found of AI-assisted discharge reconciliation measured against a pharmacist-completed reference, and the nearest studies are retrospective, single-site or synthetic¹²¹³.

Costs left out. The modelling study did not include the cost of building interoperable systems, and it noted that other prescribing errors might not fall even if transcription errors did⁴.

Over-reliance on AI output. In a preprint randomised trial of 72 physicians who had completed AI-literacy training, a prompt showing a chatbot's benchmark accuracy and a case-specific confidence signal raised scores on diagnostic vignettes by 7.6 percentage points (95% confidence interval 1.4 to 13.9). The authors cite an earlier trial in which erroneous chatbot recommendations degraded such physicians' diagnostic reasoning. The vignettes concerned diagnosis, not medicines, in a single session (randomised trial, preprint, 2026)¹⁵.

Uneven rollout of the pharmacy route. By July 2022, 27% of trusts had not started referring (trust self-reports, older than three years)⁹, and the Cornwall authors know of no formal evaluation of the service's patient benefit⁸.

Reasoning

Are the conditions realistic? The conditions that carry the most weight, an accurate admission list, shared records and a pharmacy that acts, are mostly not AI conditions. If they are not met, an AI-drafted list could be accurate and still arrive somewhere that cannot use it.

Would simpler solutions do? If a structured discharge prescription, a pharmacist check on the ward and a Discharge Medicines Service referral for every patient on high-risk medicines delivered most of the benefit, the case for AI could shrink to the cases that depend on free text. Reconciliation by people already has trial evidence behind it⁷; none was found for the AI step in this task.

Who bears the risk? The benchmark's author describes patients with many concurrent medicines as the most at risk from reconciliation errors¹³, and these were the patients for whom smaller models performed worst. If a smaller, cheaper model were chosen to save cost, the risk could fall on the people least able to absorb it.

Who would pay, and who would gain? If hospitals paid for the tool while the savings in workload and readmissions fell mainly to primary care and community pharmacy, funding could stall. The modelling study says its estimates may understate benefits in primary care⁴.

Signs to watch

• A published prospective evaluation of AI-assisted discharge reconciliation that reports missed and invented medicines against a pharmacist-completed reference. None was found, and no date for one was found.
• Medicines-specific results from live deployments of AI drafting at discharge, since the 2025 pilot reported hospital-course summaries only (search extract)¹¹.
• A formal evaluation of the patient benefit of the English Discharge Medicines Service, of which the Cornwall authors know of none⁸.
• Audits of transition medication errors before and after interoperable prescribing information is introduced, which the modelling authors called for⁴.
• Further MHRA guidance on AI software beyond ambient voice products, since the July 2026 guidance says the MHRA continues to develop guidance as the field evolves¹⁴.
• Trust-level referral, rejection and no-action rates for the Discharge Medicines Service over time; the Cornwall rejection rate was 5% in 2025 and the no-action rate 18% across the period⁸.