First things first: Health is, and will remain a human-driven domain.

Perhaps you’ve heard the phrase:
“AI will not replace doctors, but it will replace doctors who don’t use AI.”

We shall see. In the here and now, most health-related matters will have a human at the centre of each step and interaction. The reasons are operational, regulatory and emotional.

Nevertheless, we urgently need to address a severe capacity shortage and free health practitioners from tasks that AI can do, such as data acquisition, processing, and decision support.

So, when we talk about adopting AI, we mean this narrow use case: improving patient experience in the data domain, freeing up practitioners to exercise their unique judgement instead of doing data entry, and offering emotional guidance and treatment support.

Time is what patients demand the most and what busy practitioners can offer the least.

Can we bridge the gap?

Let’s define our terms first.

AI is NOT LLMs

When we say AI, we don’t mean Artificial intelligence or LLMs. We mean “Augmented intelligence”. We explain it in detail in this article, but for our purposes, let’s simply call it the intermediate step between where we are today and where we need to be once we get our data in order.

Here’s our take on where AI adoption is presently:

• The healthcare interface is changing. Text—or image-based interfaces are the future, replacing complicated dashboards or EHRs that make data retrieval and interpretation challenging. The first batch of companies introducing this “shell” user interface is already in the market, but the use cases are limited to text-based Dx and Rx, which do not integrate with current workflows but sit side by side.
• …but very slowly. Due to inertia, a lack of bandwidth to explore new solutions, and a lack of clarity of regulatory status, only a few practitioners are leveraging new tools. Insurance and the mighty EHR it is designed for still rule the data domain.
• AI commoditisation is happening much faster than we realise. When everyone has “AI”, the term becomes meaningless.
• Healthcare practitioners must differentiate through unique user experiences and data insights via novel diagnostics, distribution, and brand.
• Everyone can throw a blood test into an OpenAI chat window (not that they should), but that doesn’t mean they are leveraging the most potent AI tool.

The minimum every AI-curious practitioner should know.

Transformer tools are potent and will transform the healthcare landscape (pun intended). But like everything new, we need to understand their fundamental properties to leverage them effectively. Here’s a crash course:

“Coding” is translating what developers know into machine-executable instructions. The most important thing to understand about machine learning is that it represents a fundamentally different approach to creating software: The machine learns from examples rather than operating on an explicit set of clear rules – it interprets and learns rather than follows codified instructions.

There is an implicit problem with the machine learning approach. A lot of what we know is tacit—we can’t fully explain it. Think about teaching someone to ride a bike or recognise a familiar face. We “know” how to do it, but explaining every step in a way a machine could replicate is nearly impossible.

This idea is called Polanyi’s Paradox—we know more than we can articulate. And it’s been a significant obstacle to giving machines accurate intelligence. For decades, machines were limited to tasks we could explicitly code. This is especially true in healthcare, where tacit knowledge and learned pattern recognition by an experienced physician and multi-variate context make it very hard to extract explicit rules.

But now, machine learning (ML) is breaking through these limits. Machines learn from examples and structured feedback instead of relying solely on human-coded rules. For instance, ML models can now recognise faces—overcoming one of Polanyi’s classic paradoxes.

However, these systems are hard to interpret. Complex models, like deep neural networks with millions of connections, make decisions difficult for humans to explain fully. In other words, we’ve shifted from machines not knowing enough to knowing more than they can explain.

This brings three key risks:

Hidden Biases: Machines may unintentionally learn biases from the data they’re trained on. For example, suppose a health management algorithm is trained on a population with an above-average incidence of metabolic disease. It might unknowingly replicate those biases for frequent metabolic diagnosis and treatment without clear, explicit rules and create a skew in how it interprets objective data.

Statistical vs. Literal Truths: ML operates on probabilities, not guarantees, unlike traditional logic-based systems (deterministic, note the term). This makes it challenging to ensure a system will work flawlessly in all cases, especially in critical applications like healthcare or nuclear plant operations.

Error Diagnosis: When ML systems make mistakes, understanding why can be challenging. The complex, interconnected structures behind these models make diagnosing and correcting errors far from straightforward—especially as conditions change. Unlike databases that retrieve exact matches, LLMs generate responses based on probabilities—making them less reliable for tasks requiring precise recall.

The key takeaway: Data and its structure matter more than ever. While large language models (LLMs) may become widely available, the quality and context of the data they’re trained on will set apart genuinely effective solutions. The tools used to refine that data will constantly evolve, but having a proprietary dataset that is well structured for machine learning is the most defensible unfair advantage for a modern business.

And this brings us back to healthcare.

Why AI struggles in healthcare

Hospitals produce 50 petabytes of data annually, growing at 4x finance and media (Source: WEF, RBC). On top of this, 90+% of health data is not leveraged for AI because:

It is noisy (25% of chest X-rays are rejected), and there are 57 billion negotiated prices in US healthcare alone.
It is unstructured (80%) / or inaccessible due to integration issues – fax is still the most used healthcare integration method (Source: AWS)

It is unharmonised – multiple codes for claims/diagnosis/ontologies.

It is regulated. As a result, data acquisition requires using a straw instead of a big pipe, and most use cases have no real-time access.

Aren’t LLMs the silver bullets then? Not yet, and not really.

Why? Despite all the regulatory pressure, there needs to be a unified modern data infrastructure in healthcare. Data is siloed, unstructured, and has a multi-medium format (PDF and JPEG are the most common formats). That’s hard for AI to make sense of without a lot of additional, expensive processing.

There is also a UI problem—there is no interface where the data can be ingested/analysed. Health records are designed for insurers and admins, not for data scientists.

Further, we have an Audit problem: In healthcare, you must have an audit trail of why a decision was made if a machine suggested it. This is tricky with LLMs, as their probabilistic infrastructure means humans cannot easily understand the audit trail. A deterministic model for a repeatable process is needed.

But before we even get to the data, we have a Data input/output problem. From a patient’s perspective, they want the doctor’s undivided attention. However, due to regulations and insurance, practitioners are under pressure to organise all the data for the notes and follow-up and need to extract data from the patient quickly and succinctly. So, both parties want to receive data, but neither is keen to prioritise sending it. That’s not how effective communication works.

Finally, we also have the “42 Problem”—You must have the right question before answering it. Trolling or throwing data into models won’t work. Most organisations start with data, not the right question.

So what should I do in my practice then?

Disclaimer: No magic “AI” wizards will improve your workflow without investing time and effort. If you’re too busy to save yourself time, the process will require an explicitly earmarked time budget. This is the hurdle most of your colleagues fail at – they go too high level and “major the minors”. Instead, focus on what costs you time in data acquisition, processing, and user experience. Only then consider the specific tool.

With that out of the way, we now know that we need to prioritise the objectives by mapping out an internal process to improve rather than blindly going for “Adopting AI”.

Figuring out the use cases is the place to start. Here’s how we’d play it.

Step 1: Data acquisition and user experience.

Patient intake: If you’re still using a web form or—God forbid—paper and the resulting PDF is still consumed as a static document, you’ve got work to do there. There should be an automatic interpreter of the form to save you repetitive loops like high BMI, test for insulin resistance, recommend body fat and carb consumption reduction, and reorder blood tests in three months.
This is 30-60 minutes spent, per patient, that you must automate.
Lab tests. Ordering, interpretation, and follow-up/accountability can consume 60-75 minutes per patient. There is plenty of scope to automate this. Contact us if you need help there.

Novel diagnostics – Are you using Dexa, Genetic, Biome or body morphology tests like Styku? If so, are you still giving your clients PDFs?
If yes, that’s a problem—those PDFs contain the data key to engaging and retaining these clients, but it cannot be harnessed automatically. PDFs can only be opened, read, closed, and stored. If you can extract the markers and keep reusing them for advice and treatment and stack them like Lego bricks for your unique protocol, that’s a different story. The problem is that most EHRs don’t extract the data into individual markers, but rather store PDFs. We can help.

Lifestyle recommendations and protocols. If you have a unique set of protocols, medicine or supplement stacks, or specific templates for common conditions, how are you storing, communicating and tracking these? It’s a tedious, repetitive process, and chances are neither your administrative team or your patients are enjoying it very much. That’s a potent area to explore.

Step 2: Identify deterministic vs. probabilistic processes.

Large Language/Transformer models think from left to right, over-infer patterns, and under-infer randomness.

As a result, their use is likely to be limited to data organisation, some decision support, and eventually learning from patient data through retrieval-augmented generation (RAG).

That last one is quite a mouthful and not in scope. Soon enough, it will be.

Until regulation and tools available catch up, look for repeatable, easily auditable processes that often happen in your practice. We’ve outlined the key ones above, but specifalist practices, longevity clinics and functional medicine integrated care centres have many more of these as they handle vast amounts of valuable data.

Think about how to organise the data into a coherent system, and how to anonymise it for use in AI decision support. Note that to ensure HIPAA compliance, this step must be in place.

Step 3: Prepare your data structure for AI – say good bye to PDFs and hello to lego bricks.

We’ve made it very clear where we stand on the use of of PDFs to store and communicate data. Here’s why:

PDFs are unstructured, meaning a modern tool like a transformer model (ChatGPT or Claude) needs to ingest, interpret and parse the document into something it can use. The technicals are not important, but the allegory we would use is lego bricks – it basically goes through the box of bricks and extracts each brick so that it can know its properties. For example, for a lab test, it would extract the marker name, the value, the units and the reference range. That’s one lego brick with five properties.

Once your data is organised this way, ideally in the same database, your lab data, wearable data, imaging data, doctors’ notes, patient history, and anything else relevant to a patient have a unified structure. More importantly, your AI assistant knows what it knows and can easily retrieve the data in a consistent format, ready for your use case.

You now have a modern data structure. From here on, any query or automation can be tendered to a neural network, deterministic system or a large language model. As the tools and computation become better and faster, you’re simply hot-swapping the right tool on the same proprietary data structure.
That’s the future, not the present. But start preparing now, before the data growth overwhelms your practice. Once you have good data, there’s no limit on what you can do, as the problem you’re trying to solve will have a superior context.

The final step is prioritising what data sources you trust and your unique approach to interpreting them. But that’s the subject of another article.

A message of hope – the computers aren’t coming for us just yet

We are in the very early minutes of the AI revolution in healthcare. Yes, the AI agents are coming, but the logic behind the assistance they offer will remain opaque, initially limiting their adoption to admin work.

Tasks like data organisation, decision support rather than outright decisions, and administration of documentation and communication will be the initial frontier rather than tasks that require deeper analysis. As The Wall Street Journal puts it, “Customers will want to know why their loan application was rejected. Bank regulators will require some decisions to be explained.” The same will happen in healthcare.

Deploying AI agents is a matter of freedom from tedium. Understanding and explaining the quality of the data being used to train an AI model unlocks exceptional insights for practitioners and patients and brand-new, faster processes. This is where communication skills come in and why they’re in high demand. Human analysis—“explainability”—is the key to empowering business and enterprise AI.

Imagine a future where an AI psychotherapist takes all the case notes and puts a virtual patient on the couch, cross-referencing the case notes and suggesting areas for exploration and research for the (regulated) human doctor. Better questions, more context, and more personalisation are all desired outcomes that are within grasp.

 
Understanding an AI’s reasoning will only be half the job. The other half will be certifying a model’s mental fitness for the task at hand. No matter how sophisticated the models and systems get, humans are ultimately responsible for the outcomes of using those systems, and we don’t see this changing.

In fact, we may see the introduction of new technology increase the number of people practising healthcare, just like ATMs increased the number of bank tellers, paradoxically. In addition to helping humans process the complex body of knowledge and offering personalised training to get the required certification, we will unlock capacity in the part of healthcare that will remain uniquely human—empathy, intuition, and understanding of human irrationality.
Humans who can think on their feet, with empathy and intuition, will be better equipped to stay a step ahead of machines. This much is clear.

However, one particular aspect of healthcare makes the jobs it requires harder to replace by AI: surprises.
Surprising work involves random situations and doesn’t repeat itself every day. Everyone who’s seen shows like House or has had the misfortune of attending an ER knows how unique each case is.

Jobs full of new permutations are very hard for neural networks because they like rules and fixed boundaries, and they want to be able to do things repeatedly. They don’t like surprises.

Healthcare is ultimately a social discipline that focuses on building relationships with people. Work tasks are not about making things but making people feel things.

That makes being a doctor so rewarding – the playbook goes out the window daily. What fun!