Artificial intelligence is moving rapidly into clinical practice, while health systems continue to face gaps in guidance, governance and structured training. A recent article published in the Journal of Medical Internet Research examines hospital medicine as a setting where adoption is already under way. A linked survey found that two-thirds of hospitalists used artificial intelligence platforms in clinical work, most often large language model tools. Use was largely outside health system integration and commonly supported clinical decision-making, including differential diagnoses and management options. Hospitalists preferred medical-specific platforms over general-purpose applications. The central concern is not whether clinicians are using artificial intelligence, but whether its implementation, training, workflow design and ongoing evaluation can support meaningful gains for patients, clinicians and health systems.
Adoption Alone Does Not Define Success
Hospital medicine now faces a familiar challenge. Clinical technology can spread quickly without delivering the improvements expected from it. The experience of electronic health record implementation demonstrates that widespread use does not automatically optimise patient outcomes, health equity, cost, patient experience or clinician experience. Electronic health records have contributed to standardisation and safety, yet they have also brought interoperability problems, inefficiencies for clinicians, cognitive burden and patient safety concerns.
These problems reflect failures in design, implementation and fit rather than the full potential of the technology itself. The same risk applies to artificial intelligence tools that enter daily clinical work before organisations define how they should be used, supported and governed. Clinician-initiated adoption may move faster than system-led deployment, but both routes require deliberate implementation and clear organisational support across routine hospital care.
Three priorities follow from this context. Clinicians need training in prompt engineering and interpretation of outputs. Health systems need implementation science frameworks for deployment in clinical environments. Organisations also need strategies for evaluating effects over time. Without these elements, use rates may rise while the contribution to care quality, workforce experience and health system performance remains uncertain.
Training Shapes Clinical Value
Large language model tools can support clinical reasoning, but the value of that support depends on how clinicians interact with them. Outputs vary according to prompt input and may be overly confident or erroneous. In clinical decision-making, those characteristics create risk when users have not learned how to prompt, interpret, contextualise and act on the information generated.
Evidence from randomised controlled trials reinforces the importance of implementation and training. One trial compared diagnostic performance among clinicians alone, clinicians with artificial intelligence assistance and artificial intelligence alone. Artificial intelligence assistance did not improve clinicians’ diagnostic reasoning, while artificial intelligence alone outperformed clinicians. The result indicates a gap in how the tool entered clinical reasoning rather than a case for replacing clinicians.
Must Read: Healthcare AI Growth Outpaces Governance
Other evidence points to the importance of workflow structure. Another randomised controlled trial found that diagnostic accuracy improved when clinicians received targeted training and used intentionally designed artificial intelligence-first or artificial intelligence-second workflows. The timing and placement of artificial intelligence in clinical reasoning therefore shape outcomes. Early evidence on artificial intelligence-enabled documentation tools, including artificial intelligence scribes, also shows promise while demonstrating variable benefit across care settings and workflows. Effective adoption therefore depends on matching the technology to real clinical work, not simply adding a new platform to existing practice.
Frameworks Can Guide Deployment and Evaluation
Artificial intelligence use in clinical care requires more than deployment. Practical decisions need to address workflow integration, organisational culture, infrastructure, training, external pressures and unintended consequences. Implementation science frameworks offer a way to make those decisions systematic rather than reactive.
The pragmatic robust implementation and sustainability model, known as PRISM, provides one option for health system leaders. It can support assessment of where an artificial intelligence tool fits within workflows, which users require training, what could go wrong, how performance should be monitored and how the effects on patients, the workforce and the health system should be assessed. These questions matter because implementation quality determines whether artificial intelligence tools advance clinical and organisational goals or add new burdens.
Individual clinician responsibility also remains important. Clinicians who already use artificial intelligence tools need to understand their strengths and limitations and recognise how outputs may influence decision-making.
Evaluation cannot stop after deployment. Learning health system infrastructure can support continuous monitoring and iterative adaptation by using data generated through routine clinical care. Electronic health record data already include audit logs that reveal workflow patterns and clinical data that track patient outcomes. Measurement also needs to extend beyond easily captured data, including periodic assessment of clinician experience, workflow burden and economic value.
Artificial intelligence use in hospital medicine is likely to continue expanding, with large language model platforms already part of clinical work for many hospitalists. Its impact remains uncertain because adoption alone cannot show whether outcomes improve for patients, clinicians or health systems. Training, workflow design, implementation science and continuous evaluation are central to whether artificial intelligence supports clinical reasoning or adds risk and workload. The main challenge now lies in turning rapid uptake into careful, supported and measurable implementation across everyday hospital practice.
Source: Journal of Medical Internet Research
Image Credit: iStock
References:
Maw A, Pandita A, Burden M et al. (2026) Hospitalists Are Already Using AI—Why Implementation Will Determine Its Impact. J Med Internet Res;28:e97419.
