Medication errors remain one of the most common causes of preventable patient harm in hospitals. Despite decades of work on medication safety—barcoding, clinical decision support, pharmacist review—errors persist.

AI offers new approaches to this old problem. Some are proving genuinely useful. Others are more promise than reality. Here’s my assessment of where hospital pharmacy AI actually stands.

The Medication Safety Challenge

The medication use process has multiple steps where errors can occur:

• Prescribing (wrong drug, dose, route, frequency)
• Transcription (errors in translating orders)
• Dispensing (wrong medication or dose prepared)
• Administration (wrong patient, wrong time, wrong technique)
• Monitoring (failing to detect adverse effects or interactions)

Traditional safety systems address specific error types:

• Electronic prescribing with decision support reduces prescribing errors
• Barcoding reduces administration errors
• Pharmacist review catches errors before dispensing

AI can augment each of these, and can address error types that traditional systems don’t catch well.

AI Applications in Hospital Pharmacy

Prescription Anomaly Detection

AI that reviews prescriptions for anomalies that might indicate errors:

• Doses outside normal ranges for specific patient characteristics
• Unusual drug combinations
• Orders inconsistent with diagnosis
• Deviations from typical prescribing patterns for specific conditions

Traditional decision support uses rule-based alerts: “This dose exceeds maximum.” AI can be more nuanced: “This dose is unusual given this patient’s renal function, age, and concomitant medications.”

The advantage is catching errors that don’t violate explicit rules but are still concerning. The risk is alert fatigue if the system generates too many false positives.

Drug-Drug Interaction Prediction

Current interaction checking uses knowledge bases of known interactions. But the list of documented interactions is incomplete, and severity assessments are often crude.

AI approaches can:

• Predict interactions based on drug properties even when not explicitly documented
• Personalise interaction risk based on patient characteristics
• Prioritise interactions by likely clinical significance
• Reduce alert fatigue by suppressing low-risk alerts

Several vendors are developing these capabilities, though deployment in Australian hospitals remains limited.

Adverse Drug Event Detection

AI that monitors patient data to detect signs of adverse drug events:

• Laboratory changes suggesting drug toxicity
• Clinical deterioration patterns associated with medication adverse effects
• Documented symptoms that might indicate drug reactions

This moves from preventing errors before they happen to detecting adverse effects after administration—still valuable for minimising harm and informing future prescribing.

Dosing Optimisation

AI that recommends optimal doses based on patient characteristics:

• Population pharmacokinetic modelling to predict drug levels
• Therapeutic drug monitoring integration for drugs with narrow therapeutic windows
• Renal and hepatic dose adjustment based on real-time function assessment

For drugs like aminoglycosides, vancomycin, and anticoagulants where dosing is complex and consequences of mis-dosing are serious, AI-assisted dosing can improve outcomes.

Medication Reconciliation

AI that assists medication reconciliation—confirming that medication lists are accurate across care transitions:

• Natural language processing to extract medications from clinical notes
• Matching medications across different sources (GP records, hospital records, pharmacy records)
• Flagging discrepancies for human review

Medication reconciliation errors are common at admission and discharge. AI can reduce the manual effort while improving accuracy.

What’s Actually Deployed in Australia

In Australian hospital pharmacy practice, I see:

Basic clinical decision support: Widely deployed but often not using modern AI. Rule-based systems with significant alert fatigue.

Smart dosing for specific drugs: Some hospitals use AI-informed dosing for aminoglycosides, vancomycin, and warfarin. This is probably the most mature AI application in hospital pharmacy.

Adverse event detection: Limited deployment. Some hospitals have sepsis detection AI that integrates medication data, but dedicated adverse drug event AI is uncommon.

Prescription anomaly detection: Emerging. Some organisations are piloting but not yet widely deployed.

Medication reconciliation AI: Nascent. The technology exists but integration with Australian pharmacy systems is limited.

Barriers to Adoption

Several factors slow pharmacy AI adoption:

Integration complexity. Pharmacy AI needs to integrate with electronic medication management systems (eMM), pharmacy dispensing systems, and electronic medical records. This integration is technically challenging.

Workflow fit. Pharmacists have established workflows. AI tools that don’t fit these workflows don’t get used. Design for pharmacy workflow matters.

Alert fatigue legacy. Pharmacists already receive excessive alerts from existing systems. Adding AI alerts without reducing existing alert volume makes problems worse. AI implementation must address existing alert fatigue, not add to it.

Evidence requirements. Pharmacy leaders (rightly) want evidence that AI improves medication safety before deployment. This evidence is still developing for many applications.

Staffing constraints. Implementing and operating AI requires time from clinical informaticists and pharmacists who are already stretched.

Making Pharmacy AI Work

For organisations implementing pharmacy AI:

Start With High-Impact Applications

Focus on applications with:

• Significant patient safety impact
• Clear evidence of effectiveness
• Workflow fit with current practice
• Technical feasibility in your environment

Smart dosing for high-risk drugs is often a good starting point. Clear value proposition, established evidence, bounded scope.

Address Alert Fatigue Holistically

If implementing alert-generating AI, redesign the entire alerting environment:

• Remove or suppress low-value existing alerts
• Tier alerts by severity and actionability
• Ensure alerts are interruptive only when needed
• Monitor alert response rates and adjust

Adding AI alerts on top of existing alert overload guarantees failure.

Involve Pharmacists in Design

Pharmacist input is essential throughout:

• Identifying which problems AI should address
• Designing workflows that fit pharmacy practice
• Evaluating AI outputs for clinical relevance
• Training and change management

AI tools designed without pharmacist involvement don’t get adopted.

Measure Outcomes

Track whether AI is actually improving medication safety:

• Error rates before and after implementation
• Near-miss identification
• Clinician satisfaction with AI tools
• Alert response patterns

If you can’t demonstrate improvement, how do you know the AI is working?

Vendor Landscape

The pharmacy AI vendor landscape is less developed than radiology or pathology:

• Major eMM vendors (Cerner, Epic) are adding AI features to their products
• Specialist pharmacy AI vendors exist but may not integrate well with Australian systems
• Some local development is occurring in academic and health service partnerships

For many Australian hospitals, the path to pharmacy AI will be through their existing eMM vendor adding capabilities, rather than standalone AI products.

Organisations exploring options benefit from working with knowledgeable partners. AI consultants Sydney and health informatics consultancies can help navigate vendor options and integration approaches.

Safety Considerations

Pharmacy AI directly affects medication decisions. Safety considerations are paramount:

Validation in local context. AI trained elsewhere may not perform well with local prescribing patterns, formularies, and patient populations. Local validation matters.

Clear override processes. Clinicians must be able to override AI recommendations when clinically appropriate. Override shouldn’t be so difficult that it prevents necessary deviations.

Monitoring for errors. AI itself can introduce errors—wrong recommendations, missed interactions. Monitor for AI-caused errors, not just AI-prevented errors.

Liability clarity. When AI contributes to prescribing decisions, liability arrangements should be clear. Who is responsible when AI contributes to an error? AI consultants Melbourne recommend addressing liability questions explicitly during governance review, before implementation begins.

Looking Forward

Hospital pharmacy AI is less mature than radiology AI but has significant potential. Medication safety is a perennial problem; AI offers new approaches.

My expectations:

Next 1-2 years: Smart dosing becomes more common. Alert rationalisation projects incorporate AI. Pilot projects in prescription anomaly detection.

2-4 years: Integrated medication safety AI becomes available from major eMM vendors. Adverse event detection sees broader deployment.

5+ years: AI-augmented medication management becomes standard practice. The question shifts from “should we use AI?” to “how do we optimise AI performance?”

For pharmacy leaders, the time to start building capability is now. Even if full deployment is years away, understanding AI, building governance structures, and piloting applications positions organisations for the future.

---

Dr. Rebecca Liu is a health informatics specialist and former Chief Clinical Information Officer. She advises healthcare organisations on clinical AI strategy and implementation.