Over the next five years, AI in healthcare will most likely become routine in documentation, coding, patient messaging and imaging support, largely delivered through EHR vendors. Autonomous diagnosis at scale, AI replacing doctors, and large cost reductions are less likely. Payment rules, evidence of outcomes and liability will shape the trajectory more than model capability.
In 2023, 38% of US physicians told the American Medical Association they used AI in their practice. By early 2026, the AMA’s survey put the figure at 81%. Very few clinical technologies have moved that fast. Yet ask a hospital chief executive whether AI has changed the fundamentals of their business, such as cost per case, staffing ratios or outcomes, and the honest answer is: not yet.
That tension defines the next five years. Adoption is running far ahead of transformation. The tools are spreading quickly into the parts of healthcare that are mostly paperwork, while the parts that are mostly judgment, risk and money are moving at the pace of payment rules, regulators and courts. The central prediction of this piece is that AI’s first five years at scale will be remembered less for breakthroughs in diagnosis than for quietly absorbing the administrative work of medicine, and that whether it goes further depends on institutions, not algorithms.
What will AI in healthcare look like in five years?
Start with what is already happening, because the near future is mostly the present, spread wider.
Documentation will be largely automated. Ambient scribes went from pilots to one of the fastest-adopted technologies in healthcare in about two years. By 2030, it is likely that most clinical notes in well-resourced health systems will be drafted by AI and edited by clinicians. The debate will shift from whether to use scribes to how to audit what they write and whether shorter notes lose clinical nuance.
Revenue cycle work will be heavily automated on both sides. Coding, claims, prior authorisation and appeals are rule-bound, document-heavy and measurable, which makes them ideal for AI. Menlo Ventures estimated that coding and billing automation drew around $450 million of healthcare AI spending in 2025, second only to ambient documentation. Payers are automating too. Expect an escalating machine-to-machine contest over claims, and growing pressure on regulators to set rules for it.
Patient communication will be AI-mediated. Drafted replies to portal messages, AI assistants inside patient apps and automated scheduling and navigation will become normal. The quality question will be whether they reduce clinician workload or simply increase message volume.
Imaging AI will be standard support in radiology. The large majority of AI-enabled devices on the FDA’s list, now well over a thousand, are in radiology. Triage, measurement, and second-reader functions in screening will become routine where payment or workforce shortages justify them.
Most of this will arrive through the EHR. EHR vendors are bundling assistants, predictive models and ambient tools into their platforms. The practical future of hospital AI is less a marketplace of startups and more a set of features switched on inside the systems clinicians already use.
What is likely to be overhyped?
Three predictions deserve scepticism.
AI replacing doctors. Five years from now, AI will do more tasks that doctors used to do, especially documentation, summarisation and first-pass interpretation. But clinical responsibility, liability and patient expectations will keep humans in charge of decisions. The more realistic change is in the mix of work, not the number of clinicians. Given global workforce shortages, AI is more likely to make existing clinicians go further than to make them unnecessary.
Large, near-term cost reductions. Healthcare costs are driven by prices, chronic disease, ageing, labour and utilisation. AI can reduce administrative waste and may improve some outcomes, but it also creates new costs: licences, integration, validation and, in some cases, more testing and more treatment triggered by more findings. Some savings will be captured by whoever holds the contract, not the system. A net fall in national health spending attributable to AI by 2030 is unlikely.
General-purpose AI doctors for the public. Consumer chatbots giving medical advice are already widely used, and that will grow. But turning them into regulated, accountable services that can diagnose and prescribe at scale runs into liability, regulation and the need for physical examination and testing. Expect sophisticated triage and navigation, not autonomous primary care, in most systems.
What will actually decide the trajectory?
Four forces matter more than the next model release.
Payment. Clinical AI that has a payment route gets used; AI that does not is a cost centre. A 2023 NEJM AI analysis of US insurance claims found that only two categories of AI device, coronary assessment and diabetic retinopathy screening, had substantial billing volume, and both had dedicated codes. Unless payment systems create routes for AI-enabled services, or value-based contracts reward outcomes, clinical AI will stay concentrated in large systems that can fund it from operating budgets.
Evidence of outcomes. Most AI evidence today is about accuracy and diagnostic yield, not patient outcomes. If the next five years produce convincing outcome studies, for deterioration alerts, screening or chronic disease management, adoption of clinical AI will accelerate. If they produce mainly more accuracy papers, buyers will grow sceptical.
Regulation. Regulators are adapting, slowly. In Europe, the AI Act’s high-risk requirements will apply to AI in medical devices, but under the digital omnibus changes agreed this year, the date for AI embedded in regulated products such as medical devices has moved to August 2028. In the US, the FDA has been developing its approach to adaptive AI and AI-enabled device software. The rules written in the next few years will decide whether continuously updated models can be used clinically and who is accountable when they change.
Liability. When an AI-assisted decision harms a patient, who is responsible: the clinician, the hospital, the vendor or the model developer? Courts and legislatures have barely begun to answer. The first significant cases will shape adoption more than any product launch, because they will tell hospitals and clinicians how much they can rely on AI.
A worked example: two plausible 2030s
Consider a hypothetical mid-sized health system in 2030, under two scenarios. These are illustrations, not forecasts of any specific organisation.
In the incremental scenario, AI drafts nearly all notes and patient replies, runs much of the revenue cycle and supports radiologists. Clinicians report less after-hours work and burnout has eased. But clinical AI is patchy: a few deterioration and screening tools are validated and funded, and many others remain in pilots. Costs per case are roughly unchanged. The system’s finance team would describe AI as a necessary operating expense, like the EHR itself.
In the transformative scenario, the same administrative gains have happened, but payment reform and outcome evidence have also arrived. Deterioration models tied to response teams have measurably reduced ICU transfers. AI-supported screening has extended into primary care and pharmacies. Value-based contracts pay for avoided admissions, so the system invests in clinical AI because it earns money by keeping patients well.
The technology in both scenarios is similar. The difference is payment, evidence and governance. That is why the future of AI in healthcare is as much a policy question as a technical one.
The strongest counterargument: this underestimates the technology
The best objection to this cautious view is that it assumes today’s AI, and technology has repeatedly outrun cautious forecasts. Model capabilities have improved rapidly, and costs for a given level of performance have fallen dramatically. AI agents that can take actions across systems, not just answer questions, are already being built into EHRs. If models become reliably better than clinicians at specific diagnostic and management tasks, institutional barriers may fall faster than expected, because the moral and competitive pressure to use them will be overwhelming.
That is a real possibility, and it is the main way these predictions could be wrong. But healthcare has a long record of technologies whose capability arrived well before their institutional adoption, from electronic records to telehealth. Telehealth needed a pandemic and emergency payment rules. What would accelerate AI in the same way? A clear payment pathway, decisive outcome evidence, or a liability framework that gives clinicians confidence. Any of these would change the trajectory more than a better benchmark score.
What this means for health leaders, clinicians and policymakers
For health system leaders, plan for AI as an operating capability, not a series of projects. Build governance, measurement and integration skills now, because they are what separates organisations that benefit from those that merely spend. The argument in our predictions for healthcare in 2027 applies here: be explicit about what would prove you wrong.
For clinicians, expect your work to change more than your role. Skills in supervising AI output, spotting its errors and explaining it to patients will matter. Training matters too; we argued in teach residents to work without the AI first that core skills must be built before they are delegated.
For policymakers, the levers are payment, evidence standards and liability rules. Pay for outcomes where AI can deliver them, fund independent evaluation, and clarify who is responsible when AI-assisted care goes wrong.
For patients, AI will increasingly mediate how you reach care. It is reasonable to ask whether AI was used in your care and who reviewed it. Expect the first voice you hear when booking an appointment, and the first draft of the reply to your portal message, to come from software. The safeguard worth insisting on is a clear route to a human when the software gets it wrong.
Frequently asked questions
Will AI replace doctors in the next five years?
Very unlikely. AI will take over many tasks, including documentation, summarising records and first-pass reading of some images, but clinical responsibility, liability and patient trust will keep clinicians in charge of decisions. Given global workforce shortages, AI is more likely to extend what existing clinicians can do than to replace them.
What are the most promising uses of AI in healthcare?
In the near term, clinical documentation, coding and billing, prior authorisation, patient messaging and imaging support have the strongest adoption and clearest benefits. In the longer term, deterioration prediction, screening and chronic disease management could improve outcomes, provided they are tied to effective clinical responses and properly evaluated.
How fast are doctors adopting AI?
Quickly. The American Medical Association’s surveys found that 38% of US physicians reported using AI in 2023, 66% in 2024 and 81% in early 2026. Most uses are administrative or informational, such as documentation and summarising research, rather than autonomous clinical decisions.
Will AI reduce healthcare costs?
It may reduce some administrative costs and improve efficiency, but large national savings within five years are unlikely. AI brings new costs for licences, integration and validation, and can increase testing and treatment. Savings are more likely where payment models reward outcomes and reduced utilisation.
What could slow down AI in healthcare?
Lack of payment routes for clinical AI, thin evidence of patient outcomes, regulatory uncertainty, unresolved liability, data privacy concerns, cybersecurity risks and high-profile safety failures. Conversely, clear payment pathways, strong outcome studies and liability clarity would accelerate adoption.
The next five years of healthcare AI will not be decided in model labs. They will be decided in payment rules, outcome studies, courtrooms and hospital governance committees. The technology is moving fast. The institutions that decide what it is allowed to do are the real timeline.
What is the one change in the next five years that would most accelerate, or halt, AI in your part of healthcare? Tell us in the comments.
