A new study from CAMO-Net UK, published in The Lancet Digital Health, provides rare, real-world insight into how clinicians interact with artificial intelligence (AI) when making complex antimicrobial prescribing decisions – particularly under conditions of uncertainty.
The research evaluates an AI-driven clinical decision support system (CDSS) designed to assist clinicians in deciding when patients receiving intravenous antibiotics could safely switch to oral treatment. Early intravenous-to-oral switching is a key antimicrobial stewardship strategy, offering benefits for patients, healthcare systems, and the environment, yet it remains underused in routine practice.
Understanding AI in real clinical decision making
While many AI tools show promise in development, far fewer have been rigorously assessed in ways that reflect real clinical behaviour. This study addresses that gap by examining not only what decisions clinicians make when supported by AI, but how and why those decisions change.
Using a randomised, multimethod design, the research team engaged 42 clinicians from 23 hospitals across the UK, including consultants, trainees, pharmacists, microbiologists, and infectious disease specialists. Participants took part in:
- semistructured interviews exploring attitudes to antimicrobial prescribing and AI
- a web-based clinical vignette experiment comparing standard care with AI-supported decision making
- usability and technology acceptance questionnaires
This approach allowed the researchers to triangulate behavioural data with clinicians’ perceptions and experiences.
What the study found
Overall, the AI system was positively received and rated as usable, with a System Usability Scale score of 72 out of 100, considered “good” for healthcare software. Most participants reported that they would intend to use such a system if it were available in practice.
Importantly, the AI did not override clinical judgement. For most cases, prescribing decisions were similar whether or not the AI CDSS was present, and clinicians were able to identify and ignore incorrect AI recommendations. This finding challenges concerns that AI might lead to over-reliance or automation bias.
However, the study did observe behavioural shifts in specific scenarios. When the AI CDSS recommended not switching from intravenous to oral antibiotics – a more cautious option – clinicians were significantly more likely to follow that advice. This aligns with the broader culture of caution reported by participants during interviews, particularly in situations of uncertainty.
AI explanations: available, but rarely used
Although the CDSS provided explanations for its recommendations, these were accessed infrequently – used in only 9% of cases when available. This suggests that, at the point of care, clinicians prioritise speed, clarity, and trust over detailed technical explanations.
The findings reinforce growing evidence that explainability alone is unlikely to drive AI adoption. Instead, clinicians emphasised the importance of clinical evidence, ease of use, and seamless integration into existing workflows.
Implications for antimicrobial stewardship
Antimicrobial prescribing is inherently complex, shaped by uncertainty, patient-specific factors, and social and behavioural influences. The study highlights that AI tools must be designed with these realities in mind if they are to support stewardship effectively.
Rather than replacing clinicians, the AI CDSS functioned as a supportive, reassuring presence, particularly in borderline cases. This suggests potential value in using AI to:
- support non-specialist prescribers
- help prioritise stewardship input from pharmacists and microbiologists
- reduce unnecessary prolonged intravenous antibiotic use
However, the CAMO-Net researchers emphasise that prospective clinical trials are now needed to establish safety, patient outcomes, and system-level benefits before widespread implementation.
Dr Tim Rawson, Clinical Associate Professor in Infectious Diseases and Antimicrobial Resistance at Imperial College London, said, “This study shows that AI decision support can be acceptable and useful to clinicians without replacing professional judgement. In complex and uncertain prescribing decisions, the value of AI lies in how well it supports existing clinical workflows, builds trust through evidence, and helps clinicians navigate risk rather than directing care.”
Building evidence for responsible AI
This study represents an important step in moving AI for antimicrobial stewardship from concept to practice. By focusing on clinician behaviour, usability, and trust, rather than predictive accuracy alone, it provides a realistic picture of how AI decision support may function in everyday care.
For CAMO-Net UK, the findings reinforce the importance of carefully evaluated, evidence-based AI tools that align with clinical workflows and support, rather than disrupt, professional judgement.
