For decades, better clinical knowledge came with a high price tag.
A hospital that wanted evidence-based order sets, drug information, clinical pathways, or patient education materials could not simply create them. It usually had to license the content from a specialized publisher. Then integrate it into the electronic health record, adapt it to local workflows, maintain it, and keep paying to keep it current.
Large health systems could justify that investment. Many smaller hospitals and medical practices could not.
That created an overlooked form of inequality in healthcare. Not just who had better equipment or more specialists. Who could afford to put reliable medical knowledge inside the physician's workflow.
AI may now change that equation.
The most important impact of large language models in healthcare may not be that they can write clinical notes. It may be that they can turn clinical knowledge from a scarce, expensive product into abundant, adaptable infrastructure.
Generating clinical content is becoming nearly free. Validating it is not. The competitive advantage moves from owning knowledge to validating and delivering it.
Mexico could be one of the countries that benefits most.
Clinical knowledge became a software category
Clinical decision support is not one product. It is a pile of tools and content.
Drug interaction warnings when a physician writes a prescription. Suggested lab tests. Evidence-based order sets. Treatment pathways. Preventive-care reminders. Discharge instructions and patient education materials.
Traditionally, producing this required teams of physicians, pharmacists, editors, researchers, and software developers. A company would build and continually update a proprietary clinical library. Hospitals and software vendors would license access to it.
The model made sense. Medical knowledge is complex, changes constantly, and can cause harm when it is wrong. Publishers built real editorial and review processes around it.
But it also created barriers.
Price: the license is only part of the cost. Integration: a static library still has to be mapped to medications, diagnoses, specialties, workflows, and fields inside the EHR. Localization: content built for one healthcare system rarely matches another country's medications, language, reading level, or regulations. Scale: ten standard handouts were possible. A different explanation for every patient was not.
Large language models hit all four at once.
From a library to a generation layer
Take patient education.
Under the old model, you licensed a library of documents about diabetes, hypertension, or postoperative care. The physician or nurse picked the one that came closest.
We ran into this firsthand on a project in the United States. Even patient education, something that looks trivial from the outside, required an expensive license and a serious implementation effort. The documents existed. Turning them into something usable inside the clinical workflow was a separate software project.
An AI-enabled system can write instructions for one specific patient.
It can explain a new diabetes diagnosis in Spanish. Account for age, current medications, and lab results. Use simpler language for someone with limited health literacy. Describe warning signs, what to do next, and what can wait until the next appointment.
The marginal generation cost runs from a fraction of a cent to a few cents, depending on the model and how much context it needs. The exact number almost does not matter. The economic unit has changed from an enterprise content license to a software request.
The same shift applies across clinical decision support.
An EHR can draft an order set for a specific diagnosis, setting, and patient. Compare a proposed prescription with allergies, medications, renal function, and other known risks. Summarize the evidence behind a treatment option. Turn an institutional protocol into a checklist at the moment of care. Adapt discharge instructions to what actually happened during the encounter.
That does not mean a raw prompt should replace validated medical content. It means the system no longer needs to store every possible output in advance.
Instead of choosing one document from a library, the software can assemble a response from clinical evidence, institutional policies, and patient data when it is needed.
That is a different architecture.
Are the models reliable enough?
This argument would mean little if the models only produced convincing but medically weak text.
They are better than that, and still not enough on their own.
On some structured diagnostic benchmarks, leading models now perform at or above physician level. In one study built from 304 difficult New England Journal of Medicine cases, an AI system reached 80% diagnostic accuracy, against 20% for the participating generalist physicians. It was a controlled benchmark on unusually hard cases, not ordinary practice. Still, the frontier is moving fast.
Diagnosis is not the same task as drafting discharge instructions or an order set. But it demonstrates the capability most of those tools actually depend on: reasoning over messy clinical information.
They also fail in familiar ways. Performance drops when the test involves uncertainty, evolving information, and the ambiguity of real clinical judgment. And fluent text can still invent facts. Verification is a safety requirement, not a side effect of sounding confident.
So the useful conclusion is not "AI is already a doctor" or "AI is too unreliable for medicine."
It is that AI is capable enough to draft, retrieve, personalize, and support, if the surrounding system controls the sources, checks the output, and keeps a clinician accountable for the final decision.
Like other important medical technologies, the value is not in the component alone. It is in the system around it.
Why this matters more in Mexico
In countries where sophisticated clinical decision support is already widespread, AI may first replace or improve tools organizations already have.
Mexico starts from a different place.
A 2024 Fundación Mexicana para la Salud (FUNSALUD) study found that only about 9% of Mexican physicians were using AI tools in daily practice. In the United States, an American Medical Association survey put physician use of healthcare AI at 66% in 2024 and 81% by early 2026. The two surveys measured different things, so the figures are not directly comparable. Even so, the difference in adoption speed is hard to miss.
Normally that gap would be discouraging. Here it may be an opening.
Mexico does not need to replay every stage of clinical decision support in wealthier systems. It may not need to buy large static libraries, finish years-long integrations, and build complex rules one at a time.
It can move straight to a generation layer inside modern EHRs.
There is a useful parallel outside healthcare. In parts of Africa, mobile money leapt ahead of traditional banking. Many people had never had a branch, a checking account, or a dense card network. That absence looked like a disadvantage. It also meant there was less legacy infrastructure to defend. Systems like M-Pesa could go straight to phones because the old stack was never fully built.
Clinical knowledge in much of Mexico looks similar. The expensive licensed libraries were never widely deployed. That is a gap. It is also freedom from sunk cost.
A clinic that could never justify licensing hundreds of patient education documents may soon generate a better one for each patient. A hospital that could not buy a complete library of order sets may develop, validate, and maintain its highest-priority protocols with AI assistance. A physician on an affordable cloud EHR may get medication checks and evidence summaries that once belonged mainly to large institutions.
This is not merely a cheaper version of the existing model. It is a different way to distribute medical expertise.
The difficult part is no longer generating the content
The tempting mistake is to believe an API call solves the problem.
It does not.
Generating a plausible order set is easy. Establishing that it is safe, current, and appropriate for a particular institution is harder.
Organizations will need to know which sources the system uses. Version control. Audit trails. Clinical ownership. Regular evaluation. High-risk outputs should be constrained by approved institutional rules and trusted references. And the system has to be tested in Spanish and in Mexican clinical contexts: local drug names, availability, guidelines, and patterns of care.
They also have to decide where generation is appropriate.
A personalized explanation of how to prepare for a routine test is one risk profile. A recommendation to change an anticoagulant dose is another. The first can lean toward automation. The second needs tight controls and explicit physician review.
The future will not belong to the organization with the most impressive chatbot. It will belong to the ones that build the best governance around generated knowledge.
Healthcare software companies have a new job too. Connecting an EHR to a general-purpose model is not enough. They have to provide the context, restrictions, evidence, and workflow that make the model useful, and measure whether it improves care without creating new alert fatigue or overconfidence.
From physician tools to patient tools
The first phase will mostly support clinicians.
That is the natural starting point. Physicians can review the output, catch mistakes, and apply context the system lacks.
But it will not stop there.
Patients are already asking general-purpose AI about symptoms, lab results, and treatment options. Often outside the healthcare system, without their complete record, and without clear clinical supervision.
The answer is not to pretend patients can be kept away from these tools. It is to build safer ones.
A patient-facing assistant connected to the EHR could explain the physician's actual treatment plan rather than speculate about one. Answer questions from approved information. Flag warning signs. Route uncertain or dangerous situations to a clinician. Help someone prepare for an appointment, follow a chronic-care plan, or understand why adherence matters.
That could democratize not only clinical content, but the ability to navigate healthcare itself.
It still requires caution. People trust confident answers even when they are wrong. Patient-facing AI should not be designed as an artificial doctor. It should be designed as a controlled extension of a real care relationship.
Mexico's leapfrog opportunity
For years, the digital-health debate focused on whether Mexico could catch up with countries that had invested much more in healthcare technology.
That may be the wrong question.
When the economics of a technology change, the past matters less. Countries and institutions that did not invest heavily in the previous generation are sometimes freer to adopt the next one.
Clinical knowledge is approaching that moment.
The old model packaged knowledge into expensive, standardized libraries. The emerging model can generate it on demand, adapt it to each context, and distribute it at negligible marginal cost.
This will not eliminate medical experts. It will increase the need for them to validate sources, define boundaries, and govern how knowledge reaches clinicians and patients.
It will not eliminate established medical publishers either. Their strongest asset may no longer be the document itself, but the trusted evidence, editorial process, and validation infrastructure behind it.
What it can eliminate is the assumption that advanced clinical support must remain exclusive to institutions that can afford complex enterprise content systems.
Mexico has a chance to put high-quality clinical knowledge inside the workflows of thousands of hospitals, clinics, and practices that have never had it.
AI makes knowledge abundant. It does not make judgment abundant.