It’s become commonplace for pharmacists to use artificial intelligence to draft emails and collate information, but how well can current AI tools help formulate Home Medicines Review (HMR) recommendations?
Researchers from the University of Queensland gave five generative AI platforms (ChatGPT, Microsoft Copilot, Claude, Google Gemini and Perplexity) two written, hypothetical HMR case vignettes and asked each to draft a letter to the patient’s GP, then assessed what the tools did – and didn’t – do well.
All models identified core safety issues
Let’s start with the positives. Every tested AI tool correctly identified the obvious safety signals. For example, the models detected explicit dose limits and contraindications well.
However, they were much less reliable in choosing the best course of action or deciding which issues to prioritise.
Identifying a problem is one thing – attributing cause is another
Spotting a problem is only half the job – a useful HMR report also has to work out what’s causing it.
One HMR case featured a patient taking furosemide whose renal function had declined.
No single AI model was consistently better at working out the cause, said lead researcher Dr Wubshet Tesfaye MPS, Lecturer in Clinical Pharmacy at the University of Queensland.
‘ChatGPT and Perplexity were better in terms of presenting several possible contributors – although they also tended to use more cautious language, which sometimes leads to less specific advice,’ he said.
These two models weighed up chronic kidney disease progression, diabetes, volume status and medication effects as possible contributors to the patient’s declining kidney function – rather than identifying a single cause.
‘Copilot, Claude and particularly Gemini attributed the decline more strongly to furosemide, despite the patient’s oedema and weight gain making volume depletion less convincing,’ Dr Tesfaye said.
AI models didn’t always use the most current guidelines
The models’ recommendations to cease or switch medicines weren’t always based on current regulatory advice.
For example, all five recommended ceasing ranitidine based on 2020 regulatory action, but none picked up that the Therapeutic Goods Administration revoked the suspension for some ranitidine products in 2024.
Dr Tesfaye said the fact that all five models missed the more recent regulatory update, suggests older, frequently cited information can dominate – unless a model reliably retrieves and applies current sources.
But it’s important to note that these findings only apply to the free/default versions of the AI models, and should not be generalised to all AI models.
‘Paid versions may provide access to capable models, stronger reasoning and better, more recent information – and could potentially have performed differently,’ he said.
The research team is now comparing outputs between free and paid models to determine whether subscription level makes a meaningful difference.
‘For pharmacists, recent regulatory changes must still be independently verified regardless of the model or subscription level used,’ he cautioned.
The patient’s individual circumstances and preferences were missing
A good HMR report should reflect the patient’s day-to-day life, not just their medicines list.
While the AI tools were able to incorporate some patient social factors, specific patient factors such as routines, hobbies and regimen burden were overlooked – unless specifically linked to an identified problem.
The models also overlooked a lifestyle-first approach to managing one patient’s diabetes, as well as a formal cardiovascular risk assessment.
The tools were most patient-centred when the relevant information was spelled out in the prompt and linked directly to a practical intervention, Dr Tesfaye said.
‘This clearly highlights the importance of obtaining such information from patients and contextualising them when prompting those tools,’ he said.
Does AI actually save time on HMRs?
It’s too early to say. The study didn’t directly measure time, so a net efficiency benefit can’t yet be claimed.
‘While it was clear that AI may save time by organising case information, flagging potential problems, and producing an initial report structure – that benefit could be reduced, or even lost, if pharmacists must correct lengthy, repetitive or inaccurate outputs and verify every dose, interaction and recommendation,’ Dr Tesfaye said.
He said the main risks include:
- outdated or unsupported advice
- poor prioritisation
- failure to integrate the patient’s full context
- automation bias
- privacy concerns
- unclear accountability.
‘The appropriate model is therefore a “human-in-the-loop” approach where the pharmacist verifies, contextualises and is responsible for the final report,’ Dr Tesfaye added.
Top tips for pharmacists undertaking medicine reviews
Based on the results of this project, the authors proposed an implementation model for use of AI in preparation of reports for medicine reviews:
- Use AI for a first-pass report: AI may be useful in drafting reports and flagging key issues.
- Verify everything: an essential step where the pharmacist checks every dose, interaction and priority against trusted current sources, such as the Australian Medicines Handbook or Therapeutic Guidelines.
- Integrate patient context: including application of kidney function, patient priorities, adherence barriers and a transparent risk matrix.
- Own the final report: approve the GP recommendations and remain accountable.
Testing AI-assisted HMRs at scale
The two cases used in this study were an exploratory proof of concept, Dr Tesfaye said.
‘The next step is to test a larger and more diverse set of HMR cases, including different clinical complexities and patient populations, with repeated runs to assess consistency,’ he said. ‘Outputs should be reviewed blindly by multiple credentialed pharmacists against an explicit benchmark.’
One of Dr Tesfaye’s PhD students is developing a study comparing AI-assisted and usual HMR workflows. The study will measure how long it takes to produce a safe, finalised report, as well as accuracy, prioritisation, unsafe omissions, correction burden and usability. Only then would the tools be tested on real HMR data and patient outcomes.
‘We are also currently running a survey and interview with credentialed pharmacists to understand the extent and nature of use of these tools in practice, which will help us inform this next step,’ he added.
This research was presented by Dr Wubshet Tesfaye as a poster presentation at the 2026 FIP Congress in Montreal on 31 August.
