The healthcare dog must wag the AI tail

Artificial intelligence can strengthen health care, but only when we focus on patient care rather than the shiny new AI models.

Artificial intelligence is changing health care at remarkable speed. Hospitals can now draw on near-real-time data to detect deterioration, flag possible sepsis, anticipate demand and identify medication risks. Generative AI can summarise notes, draft discharge information and support clinical documentation. These are genuine advances. Yet a harder question remains: has AI changed health care itself, or merely given us faster tools with potential for improvement?

My assessment is that AI has altered the means of improvement more than the outcomes. It is helping health services move from periodic measurement towards continuous sensing, and from retrospective dashboards towards prediction and decision support. But continuous sensing does not automatically produce continuous improvement, and prediction is not prevention.

An alert only helps when somebody can act on it. A deterioration score needs a clear response pathway, available staff, agreed escalation criteria and accountability. Without these foundations, more alerts can mean more noise, workload and fatigue. A technically accurate model that does not fit clinical workflow may be inert; at worst, it may distract clinicians or create false reassurance.

The same principle applies to learning health systems. Data can reveal variation and feed lessons back into practice, but software alone does not create a system that learns. That requires leaders who support transparency, teams with time and skills to improve care, useful feedback loops and a culture that can examine failure without blame. AI tends to amplify the strengths and weaknesses already present in an organisation.

It also brings new risks. Algorithms can perform differently across populations, become less reliable as clinical practice changes, or reproduce inequities hidden in historical data. Generative systems can produce fluent but incorrect answers. Clinicians may over-rely on automated recommendations, while patients may reasonably wonder who is accountable when the technology is wrong.

Governance is therefore not an administrative extra. Every AI-enabled service should have a named owner, documented purpose, defined response pathway and measures of benefit and harm. Performance should be tested across relevant patient groups and monitored after deployment for bias, errors and drift. Staff need enough practical AI literacy to understand what a system does, what it does not do and when its output should be questioned or overridden.

Consider stroke triage or sepsis warning systems. AI may identify a time-critical patient earlier, but benefit follows only if imaging is available, the right team receives the alert, transfer arrangements work and treatment begins promptly. False positives can increase workload; false negatives can encourage misplaced confidence. The important measure is therefore not simply whether the algorithm predicts accurately in a test dataset. It is whether the whole service responds more quickly and appropriately, produces better patient outcomes, avoids preventable harm and does so fairly across the population it serves.

Health services should also evaluate AI with the same discipline expected of other clinical interventions. Where feasible, we should compare outcomes before and after implementation, use control groups or staged rollouts, and specify in advance what success will look like. Models that do not deliver net benefit should be changed or retired. The aim is not to accumulate impressive AI pilots, but to improve care reliably and at scale.

For hospital leaders, the near-term agenda is practical: choose high-value problems; design the workflow and response system before deployment; involve clinicians, patients, improvement specialists and technical teams from the outset; make equity a measured outcome; and invest in implementation capability and education.

AI is a powerful ingredient, not the recipe. Safer, fairer and more reliable care will still depend on people, organisational learning and disciplined redesign. The healthcare dog must wag the AI tail, and not the other way around.

^{Based on: Braithwaite J. 2026. Artificial intelligence, data science and healthcare improvement: what has actually changed and has it changed enough? BMJ Innovations. Published online 17 August 2026. doi:10.1136/bmjinnov-2026-001608}

*Professor Jeffrey Braithwaite is Founding Director of the Australian Institute of Health Innovation at Macquarie University and Director of the Centre for Healthcare Resilience and Implementation Science. He is Chair of the International Academy of Quality and Safety.

Top image credit: iStock.com/Drazen Zigic