Brazil has the largest healthtech market in Latin America. Hospitals with global excellence standards, some of the most qualified healthcare professionals and, in the last three years, an unprecedented volume of investment in Artificial Intelligence in the sector’s history.

And yet, the vast majority of AI projects in healthcare do not deliver what they promised.

Not due to algorithm failure or lack of technical talent. The problem is earlier and more fundamental than most discussions about AI are willing to admit: the system that should support this intelligence is simply not prepared for it.

This is not an empirical hypothesis, it is verified data. The State of AI in Business 2025 report, produced by MIT NANDA, based on interviews with 52 organizations and analysis of over 300 AI initiatives, reached a conclusion that should reconfigure the entire conversation about digital transformation in healthcare: 95% of generative AI projects in companies fail or do not advance beyond the pilot. Only 5% of integrated implementations extract measurable value in terms of operational and financial results.

These numbers do not reflect limitations of the models, nor of regulations, nor of the technical capacity of the teams. According to MIT itself, what separates the 5% that succeed from the 95% that get stuck is, first and foremost, a matter of approach and, more specifically, of infrastructure.

What is behind the screen

When an institution adopts AI to automate a customer service process, predict no-shows, optimize medical scheduling, or suggest clinical protocols, the AI model does not operate in a vacuum. It depends on a foundation that rarely receives the attention it deserves: business rules structured in the system, clean data at the entry point, systems that integrate and exchange information in real time, and digitized processes that actually record what happens in the operation.

In practice, most healthcare operations have deficiencies in this foundation. Business rules, especially in complex processes, are applied outside the system, the involved systems do not integrate completely and securely, the data is not sanitized, and the traceability of actions is full of holes.

In this scenario, AI does not transform: it amplifies the noise, inconsistent data grows at scale, errors that were once human and isolated become systemic. Artificial intelligence, without the proper infrastructure to support it, is not a solution, it is a multiplier of already existing problems.

The system needs to understand healthcare, not just data

Here is the point that rarely appears in digital transformation projects in healthcare: it is not enough to integrate systems, the underlying architecture needs to be smart enough to understand the specific rules of the sector.

Which exams have conflicting preparations? What restrictions apply by health insurance, patient’s clinical profile, or service hours? Which protocols vary by unit or specialty?

When these rules exist only in the tacit memory of employees, in unstructured and incomplete textual guidelines, or in parallel spreadsheets outside the system, any automation attempt hits a critical and insurmountable limit: you cannot automate what depends on implicit judgment. The AI will operate with partial vision, and the mistakes it makes will not be small or isolated.

Well-made connections mitigate errors, and fragile connections multiply them.

Each breaking point in this chain is an opportunity for error, and when an AI operates on this fragmented structure, it does not correct errors, it scales them.

Therefore, the most relevant question today is not which AI to adopt. It is a prior question: is our system robust enough to support intelligence?

Why the market does not talk about this

The short answer is: because it doesn’t sell. The AI narrative is seductive, with promised ROI, flawless demos, and cases in large foreign institutions that do not necessarily apply automatically to the local context.

The infrastructure narrative is less glamorous; talking about structuring business rules, system integration, data quality, and process digitalization does not make headlines, but it is what separates projects that become references from those that become canceled lines in the budget.

The four gaps that no one maps before the project

Over more than 25 years working with the largest healthcare operations in Brazil, we have identified four gaps that consistently appear as obstacles before any AI can work:

GAP 1. Business rules outside the system

In healthcare, there are thousands of rules that need to be applied even in the most routine actions, such as scheduling exams. What seems simple to the layperson’s eye is actually an extremely complex process, requiring precise coordination of several variables, with very low fault tolerance. Unstructured, off-system rules generate obvious risks, which are typically addressed with a combination of employee training, layers of supervision, ombudsman channels, and good luck charms. In this context, AI, no matter how good it is, will perform an operation with similar risks, requiring even more supervision, and inevitably its use will be limited to simpler scenarios, frustrating original expectations.

GAP 2. Disconnection between systems

Poor integrations are old villains. With AI, these same villains that harmed so many initiatives, burdened processes, and generated dissatisfaction in the past, will not disappear. On the contrary, the competitive pressure to deploy disruptive AI solutions makes these issues even more relevant, as new AI agents act upon existing systems in processes that often require the correct coordination among several of them.

GAP 3. Unstructured, non-standardized, and uncleaned data

AI is impressive for its power to handle unstructured information. And rightly so, it is a revolutionary technology with potential that still knows no boundaries. But this admiration can give a false sense of security where it does not exist. Unstructured, incomplete, or simply wrong data is swampy ground on which to build AI buildings. The false idea that “you just have to tell AI what to do” and everything works is very dangerous.

GAP 4. Patient Experience not digitized from end to end

In the digital Patient Experience, it is common to have processes that still essentially depend on human intervention because they present complexities that the patient cannot handle, preventing complete digitization. Each of these not yet digitized stages represents a warning sign – if the stage has not yet been digitized, there is a reason. Despite the great power that AI technologies offer today, this reason may still be beyond AI’s reach. Often, they are structural reasons, reflecting limitations of the processes and systems themselves that should be able to support AI solutions.

These four gaps are not AI problems, they are infrastructure problems, and no model, no matter how sophisticated, solves what happens before it enters the scene if it does not have the intelligence and robustness to operate.

The sequence that works

The correct order is not: choose the AI, buy the platform, and implement it.

The sequence that generates results is: assess and address the gaps above, and then build an AI adoption plan tuned to the macro plan of technological evolution.

This sequence seems obvious on paper, but it is systematically ignored because the pressure for innovation is always greater than the pressure for foundation. And because fundamentals do not appear in demos.

Infrastructure precedes intelligence

That AI will transform healthcare is not up for debate. But the path is not so obvious and easy. MIT is one of the institutes that has invested the most in proving the potential of AI, but the same researchers who document this potential are the ones who identified, with methodological clarity, why 95% of initiatives do not get there: because organizations skip the step that would make the rest possible.

The speed and depth of digital transformation in healthcare depend less on the models that will be adopted and more on the infrastructure that supports them. Organizations that understand this sequence do not chase trends; they create the foundation to evolve predictably, incorporating new technologies with real return, without empty kickoff promises.

Infrastructure is not the opposite of innovation; it is what makes it possible.

At Touch Health, we build the infrastructure that makes digital healthcare intelligence truly work: systems that understand healthcare rules, integrate with precision, have enough robustness to support intelligence, and reduce errors before they happen.

If you want to understand how this applies to your operation, talk to our specialists:
https://linktr.ee/touchhealthtec