Artificial intelligence is entering hospitals, clinics, and clinical workflows. But between the promises of automation and the realities of patient care, there's a more interesting question: What happens when we actually put it to work?
If you've spent enough time working in a hospital, you know that healthcare is an extraordinarily complicated system.
A single patient encounter can involve physicians, nurses, pharmacists, respiratory therapists, laboratory technicians, radiologists, case managers, and dozens of other people. Each contributes a small piece of information or expertise toward a shared objective: figuring out what's wrong and deciding what to do about it.
Somehow, all of that information has to move between people, departments, software systems, and shifts.
And it doesn't always move particularly well.
We have electronic health records, automated medication dispensing systems, continuous physiological monitoring, clinical decision support, and increasingly sophisticated diagnostic equipment. Yet clinicians still spend enormous amounts of time documenting information, searching for information, reconciling conflicting information, and communicating information that already exists somewhere else.
Now artificial intelligence is entering that environment.
Depending on whom you ask, AI is either about to revolutionize medicine or introduce an entirely new category of problems into an already overburdened healthcare system.
I suspect the reality will be considerably more interesting than either prediction.
AI Isn't Coming to Healthcare. It's Already Here.
When people talk about artificial intelligence in medicine, they often imagine a future in which an AI examines a patient, makes a diagnosis, and recommends treatment.
That's an understandable image, but it doesn't accurately represent most of what AI is doing in healthcare today.
The actual applications are considerably more specific.
AI systems are helping clinicians document encounters, analyze medical images, identify patterns in physiological data, prioritize certain findings, and automate administrative work.
These applications don't all use the same technology. Some rely on traditional machine learning, others on deep neural networks, and newer systems increasingly incorporate large language models.
Their capabilities, limitations, and regulatory requirements vary substantially.
But they share something important: they're being inserted into workflows that previously depended almost entirely on human attention.
And that's where things get interesting.
The Documentation Problem
Let's start with something almost every clinician can appreciate.
Documentation.
Modern healthcare generates an enormous amount of it. Notes, assessments, medication reconciliations, care plans, discharge summaries, billing information, and countless other records.
Much of this documentation serves an important purpose. But the amount of time required to produce it can be difficult to justify, particularly when it competes with time spent directly caring for patients.
One of the more visible applications of generative AI in healthcare is ambient clinical documentation.
These systems listen to patient-clinician conversations and generate draft notes that a clinician can review, modify, and sign.
Organizations including Mass General Brigham have described using this technology in clinical encounters.
The idea is straightforward: instead of requiring a clinician to manually translate an entire conversation into structured documentation, software performs the initial transcription and summarization.
That sounds useful.
And early research suggests it can be, although the results are more nuanced than the marketing.
A randomized trial evaluating two ambient AI documentation tools found that one produced a modest reduction in documentation time, while the other did not demonstrate the same statistically significant improvement.
That's an important distinction.
A technology can be impressive without necessarily producing a meaningful improvement in the workflow it was designed to fix.
There's also a second consideration.
Someone still has to verify the documentation.
An AI-generated note can omit important details, misinterpret statements, or include information that wasn't actually discussed. The clinician remains responsible for ensuring the final record is accurate.
So the real question isn't simply whether AI can write a note.
It's whether generating, reviewing, and correcting that note requires less effort than documenting the encounter without AI.
That distinction will become increasingly important as these systems spread.
Seeing Patterns Humans Might Miss
Another major application involves medical imaging and physiological data.
Healthcare already produces far more data than any individual clinician could reasonably examine in isolation.
A hospitalized patient might generate thousands of physiological measurements during a single admission. Imaging studies produce enormous amounts of visual information. Laboratory results accumulate over time.
AI is particularly well suited to certain forms of pattern recognition within these datasets.
The FDA maintains a public list of AI-enabled medical devices authorized for marketing in the United States. Applications include medical imaging, cardiovascular analysis, and other specialized clinical functions.
Some systems help identify suspicious findings in medical images. Others assist with image reconstruction, measurement, or prioritization.
AI-based diabetic retinopathy screening is another example of a focused application that can evaluate retinal images for signs of disease.
These tools are not equivalent to a general-purpose AI physician.
They are designed to perform particular tasks under particular conditions.
That specificity matters.
An algorithm trained to identify a particular abnormality in a particular type of image cannot automatically be assumed to understand the patient's broader clinical situation.
It may identify a pattern.
But identifying a pattern and understanding what that pattern means for an individual patient are not always the same thing.
The ICU Is an Interesting Test Case
Critical care offers an especially interesting environment for thinking about artificial intelligence.
Consider a typical ICU patient.
There may be continuous electrocardiographic monitoring, invasive blood pressure measurements, oxygen saturation, ventilator data, laboratory trends, medication infusions, intake and output measurements, and a constantly evolving clinical assessment.
Each piece of information has meaning.
But its significance often depends on the others.
A blood pressure of 85/50 might be alarming in one patient and relatively expected in another.
A rising heart rate could reflect pain, hypovolemia, fever, anxiety, medication effects, or a developing physiological problem.
An elevated lactate may be important, but it cannot be interpreted responsibly without considering the broader clinical context.
Experienced clinicians don't simply respond to isolated numbers.
They integrate information.
They look for trajectories, relationships, inconsistencies, and changes from baseline.
AI may eventually become extraordinarily good at recognizing some of these patterns, particularly when it can analyze more information simultaneously than a human can practically track.
But there's a problem.
A clinically useful alert is not the same thing as a statistically accurate prediction.
Suppose an algorithm predicts that a patient has an increased probability of deteriorating within the next six hours.
What happens next?
Does it notify the nurse? The physician? The rapid response team?
What information does it provide to justify the warning?
Does the alert identify something actionable, or does it simply tell the care team that a patient they already know is critically ill remains critically ill?
And what happens when the system is wrong?
Hospitals already struggle with alarm fatigue. Adding another source of notifications without improving the decisions clinicians can make could actually make the environment worse.
The value of predictive AI depends not only on whether it recognizes a problem, but also on whether that recognition improves patient care.
The Problem With Being Almost Right
One of the more subtle dangers of artificial intelligence is that its outputs can look convincing even when they're incorrect.
This is especially relevant to generative AI.
A language model can produce a beautifully organized explanation, a plausible clinical summary, or a confident recommendation without necessarily having established that every underlying statement is accurate.
In ordinary applications, that might be inconvenient.
In healthcare, it can be dangerous.
Imagine an AI-generated handoff summary that correctly captures nearly everything about a patient but omits a recently documented medication reaction.
Or a system that summarizes a patient's history while incorrectly presenting a discontinued medication as active.
These are not necessarily dramatic failures.
They're small errors embedded inside otherwise useful outputs.
And that makes them particularly difficult to recognize.
Humans are susceptible to automation bias: the tendency to place excessive trust in information or recommendations generated by automated systems.
As AI becomes more accurate, this problem may become more complicated rather than less.
A system that's wrong half the time is easy to distrust.
A system that's correct 99% of the time may be much harder to question during the remaining 1%.
In clinical environments, those exceptions matter.
The challenge isn't merely teaching AI to generate better answers.
It's designing systems that make uncertainty visible, preserve the underlying evidence, and encourage appropriate verification.
Does AI Actually Reduce Work?
There's a recurring pattern in healthcare technology.
A new system is introduced to simplify a process.
It successfully automates part of that process.
But it also creates new tasks.
Someone has to configure it, monitor it, correct its errors, document its use, and troubleshoot its failures.
Electronic health records illustrate how complicated this can become. Digitizing information made it more accessible in many ways, but it also created substantial documentation and workflow burdens.
AI could follow a similar trajectory.
A system that automatically generates clinical summaries might save time.
But if those summaries require extensive verification, or if clinicians must navigate several additional screens to review them, the improvement may be smaller than expected.
The same principle applies to predictive alerts, automated coding, clinical decision support, and virtually every other application.
We should evaluate these systems based on what happens to the entire workflow, not just the portion being automated.
Does the clinician spend less time interacting with software?
Does the patient receive better care?
Are important errors detected earlier?
Does the system improve communication?
Are clinicians able to spend more time on tasks that actually require their expertise?
And perhaps most importantly, does it make healthcare safer?
Those are considerably more meaningful questions than whether an AI can complete a task in a demonstration.
The Next Step: From AI Tools to AI Systems
So far, most of this discussion has focused on AI performing relatively narrow tasks.
But the technology is moving toward something more ambitious.
Instead of asking a single model to generate a note or analyze a dataset, developers are beginning to build systems in which multiple specialized AI components coordinate to complete larger objectives.
This is often described as multi-agent orchestration.
The concept is relatively simple.
Different agents are assigned different responsibilities. One retrieves information. Another analyzes it. Another evaluates the result. A coordinating system determines how their work fits together.
Think of it less like asking one AI to know everything and more like organizing a team of specialized software workers.
In healthcare, the theoretical possibilities are fascinating.
Consider a future clinical system designed to help clinicians prepare for morning rounds.
One agent retrieves overnight events from the electronic health record.
Another reviews relevant laboratory trends.
A third examines medication changes and potential interactions.
Another summarizes recent imaging reports.
A coordinating agent assembles these findings into a concise clinical overview, preserving links to the underlying records.
A separate verification component checks whether the summary accurately represents the source information and flags discrepancies.
The clinician then reviews the result.
Instead of manually searching through multiple sections of the record to reconstruct what happened overnight, the clinician receives a structured starting point.
That's a much more interesting application than simply asking a chatbot to summarize a chart.
And importantly, it doesn't require the AI to independently practice medicine.
It requires the AI to organize information effectively, expose its reasoning and evidence where possible, and help a qualified clinician make better-informed decisions.
This is a hypothetical architecture, not a claim that such a fully integrated, validated system is already routine clinical practice.
What If AI Could Coordinate the Hospital?
Let's extend that idea further.
Hospitals aren't just collections of patients.
They're complex networks of interdependent processes.
A patient's discharge may depend on medication reconciliation, transportation, case management, equipment delivery, physician orders, insurance authorization, and patient education.
A delay in one part of that process can affect everything downstream.
Meanwhile, another patient may be waiting in the emergency department for the bed that hasn't become available.
An AI orchestration system could theoretically track these dependencies and help coordinate the people and resources involved.
It might recognize that a patient is approaching discharge readiness, identify outstanding administrative requirements, and notify the appropriate team members before those requirements become bottlenecks.
It could help prioritize tasks, surface missing information, and reduce repetitive communication.
Of course, there are substantial obstacles.
Healthcare information is fragmented across systems that don't always communicate well. Patient privacy requirements are strict. Clinical workflows vary between institutions. Real-world conditions are messy.
And hospital operations involve human decisions, competing priorities, and resource constraints that cannot simply be optimized away.
Still, the underlying opportunity is compelling.
Some of the most valuable applications of AI in healthcare may have less to do with replacing clinical judgment and more to do with reducing the enormous coordination burden surrounding it.
Who Checks the AI's Work?
This brings us to what I think may become one of the most important questions in healthcare AI.
As these systems become more autonomous, who is responsible for verifying what they do?
Imagine an AI agent retrieves the wrong laboratory result.
A second agent interprets that result correctly.
A third generates a recommendation based on the interpretation.
A fourth produces a polished summary for the clinician.
Every downstream component might function exactly as designed.
Yet the final output would still be wrong because the original information was incorrect.
This is a fundamental challenge of coordinated AI systems.
Errors can propagate.
Adding more agents doesn't automatically make a system more reliable. In some cases, it creates more opportunities for failure.
A verification agent can help, but only if its verification is genuinely independent and grounded in trustworthy evidence.
Otherwise, we risk building systems in which one AI confidently agrees with another AI's mistake.
Healthcare applications will need careful attention to data provenance, audit trails, permissions, uncertainty, and human oversight.
It should be possible to determine where a piece of information came from, what transformations occurred, and which components contributed to a final output.
And some decisions should remain outside the authority of autonomous software altogether.
The objective should not be maximum autonomy.
It should be appropriate autonomy, with accountability proportional to the consequences of failure.
Where Could We Be in Ten Years?
Predicting technology ten years into the future is notoriously unreliable.
Still, I think several developments are plausible.
First, AI will likely become increasingly invisible.
Instead of opening a separate AI application, clinicians may interact with intelligent capabilities embedded directly into the systems they already use.
Documentation, information retrieval, chart review, scheduling, and administrative coordination could increasingly happen in the background.
Second, clinical AI may become more context-aware.
Today's systems frequently operate within relatively narrow boundaries. Future systems may combine longitudinal patient records, physiological trends, imaging findings, medication histories, and other relevant information into more comprehensive representations of a patient's condition.
Whether those capabilities translate into safer clinical decisions will depend on rigorous evaluation, not just technical sophistication.
Third, healthcare organizations may adopt more coordinated AI architectures.
Specialized agents could handle distinct information-processing tasks while centralized systems manage permissions, verification, and workflow integration.
That could make AI considerably more useful than a collection of disconnected tools.
Fourth, the clinician's relationship with information may change.
Rather than spending so much time finding, organizing, and documenting information, clinicians could increasingly focus on evaluating it, contextualizing it, and deciding what it means for the person in front of them.
At least, that's the optimistic version.
There's also a less appealing possibility.
We could create an environment filled with automated alerts, questionable summaries, fragmented AI products, and additional verification requirements.
A hospital where everyone spends less time typing but more time checking whether software did something wrong isn't necessarily an improvement.
The direction we take will depend as much on implementation, incentives, governance, and clinical involvement as on the underlying technology.
The Part AI Can't Easily Replace
Working in critical care teaches you that healthcare is not simply an information-processing problem.
Information matters enormously.
But so do experience, communication, trust, timing, and judgment.
Sometimes the most important clinical observation is a subtle change that doesn't fit neatly into a structured field.
Sometimes a patient's condition looks acceptable on a monitor, but something about their presentation makes you uncomfortable.
Sometimes the right decision depends on understanding a patient's goals, their family, or circumstances that aren't fully represented in the medical record.
AI may become increasingly capable of helping clinicians recognize patterns and understand complex situations.
That doesn't mean we should assume it can reproduce every dimension of clinical care.
And it certainly doesn't mean that greater computational capability automatically produces better patient outcomes.
The technology should be evaluated according to what it helps people accomplish, not how closely it resembles a human professional.
The Future Isn't an AI Doctor. It's a Better Healthcare System.
I don't think the most interesting question is whether artificial intelligence will eventually become better than physicians or nurses at particular tasks.
For some narrowly defined tasks, it already performs remarkably well.
The more important question is what happens when those capabilities are integrated into the larger healthcare system.
Can we reduce the time clinicians spend searching through records?
Can we recognize deterioration earlier without overwhelming staff with alarms?
Can we coordinate care more effectively?
Can we make critical information easier to find, verify, and act upon?
Can we remove some of the administrative friction that makes modern healthcare so exhausting?
And can we accomplish those things without introducing new risks that outweigh the benefits?
Those are difficult engineering problems.
They're also clinical problems, organizational problems, and human problems.
Solving them will require more than powerful language models. It will require careful system design, reliable integrations, meaningful validation, and ongoing collaboration with the people who actually deliver care.
Artificial intelligence is already in healthcare.
What it becomes next is still very much an open question.
Personally, I'm less interested in the idea of AI replacing the people working inside hospitals.
I'm far more interested in whether we can use it to build hospitals that work better for everyone.
Sources and further reading
- U.S. Food and Drug Administration: "Artificial Intelligence-Enabled Medical Devices" (https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices) — Public information about AI-enabled medical devices authorized for marketing.
- Mass General Brigham: "Official website" (https://www.massgeneralbrigham.org/) — Institutional information on clinical AI initiatives and ambient documentation.
- World Health Organization: "Ethics and Governance of Artificial Intelligence for Health" (https://www.who.int/publications/i/item/9789240029200) — Guidance on safety, accountability, and human oversight.
- National Institute of Standards and Technology: "AI Risk Management Framework" (https://www.nist.gov/itl/ai-risk-management-framework) — A framework for identifying and managing risks associated with AI systems.
The future scenarios discussed in this article are speculative. Current AI tools vary in their clinical validation, authorization, capabilities, and intended uses. This article is commentary and is not medical advice.