.png)
GB
By Dr Declan Kelly, Founder & CEO, Eolas Medical Reading time: ~10 minutes
Every few years in healthcare, a new technology arrives that genuinely changes how clinicians work. Not the hype cycles, not the demos at conferences, not the press releases, the actual technologies that get used, every day, on every shift, by ordinary clinicians doing ordinary work.
We've had four of these waves in clinical AI. Three are well underway. The fourth is only just beginning, and it's the one that will matter most.
This is a piece about that fourth wave. About why it has taken longer to arrive than the others. About why it's structurally harder, and structurally more defensible. And about why every health system leader we speak to is now starting to explore the same question: how do we use AI on all our internal, context specific knowledge.
The first wave of clinical AI was diagnostic imaging. It is now “boringly” well-established, which is the highest compliment a clinical technology can receive.
The wedge product was AI for stroke detection. A convolutional neural network reads a non-contrast CT in 1–6 minutes, flags suspected large vessel occlusions, and pages the stroke team in parallel with the radiologist. The companies that won, e.g.. Viz.ai, RapidAI, Aidoc, Brainomix built their distribution on a simple promise: door-to-needle time, measurably reduced.
As of early 2026, 13+ AI stroke detection systems have FDA clearance, and approximately 20–30% of US hospitals with stroke programmes have adopted one of them, with concentration in larger urban centres.
On the diffusion curve, AI in imaging has comfortably crossed the chasm. It is being adopted by the late majority. The conversation has shifted from "should we use AI for this" to "which vendor, which workflow, how do we pay for this?" That is what category maturity looks like.
It took roughly fifteen years from the first deep-learning radiology papers to get here.
The second wave was ambient voice. The wedge was the consultation itself, a mic sitting in the room, capturing the doctor-patient conversation, and turning it into a structured clinical note in under a minute.
What unlocked Wave 2 was not better microphones, it was Large Language Models (LLM). These note taking/scribe/dictation platforms did exist for years prior to LLMs, they were just a bit sh*t.
For the first time, you could go from raw transcript to clinical-grade note, with “adequate” accuracy and appropriate format, without a human in the loop. This has changed the entire economics of clinical documentation and is now unlocking significant downstream opportunities, i.e. taking actions based on that captured note and giving automated clinical decision support based on that note..
The adoption numbers tell the story. Microsoft's DAX Copilot is deployed in over 600 healthcare organizations. Abridge has signed more than 200 US health systems including Mayo Clinic, Duke, Johns Hopkins, UPMC and Kaiser Permanente. When a health system makes an ambient AI scribe widely available, adoption among offered clinicians typically runs at 20–50%, with best-in-class deployments hitting 75–80%. Burnout is measurably lower. After-hours documentation time is measurably shorter. The randomised trial evidence is now in.
And remember (controversial take) evidence is a lagging indicator of the value of these technologies.
Ambient AI sits firmly in the early majority. It is no longer a question of whether a major health system will deploy ambient AI; it is a question of which one, how to pay for it etc. The companies that win Wave 2 will earn the right to be the first AI surface every clinician touches in their day, a strategically enormous position that appeared at least initially to be the only technology in decades to pose a threat to EHRs themselves.
The third wave is external clinical knowledge i.e. AI answer engines built on top of the world's published medical literature.
Wave 3 exists because of one structural problem in medicine: the volume of new clinical evidence is now growing faster than any human being can possibly read it. Medical knowledge is widely estimated to double approximately every 73 days. For a clinician to stay current with a single sub-specialty would require something like nine hours of reading per day, every day. That is not a workflow problem. It is a physics problem. No clinician can solve it on their own.
AI answer engines do solve it. The wedge product that achieved escape velocity was OpenEvidence, the clinician-facing search interface that takes a natural-language clinical question, retrieves from 35 million peer-reviewed papers and clinical guidelines, and returns a cited, synthesised answer in seconds.
OpenEvidence has reached more than 50% of US physicians inside three years and raised at a $12 billion valuation in early 2026, and crossed a critical threshold in spring 2026 when Sutter Health, Mount Sinai (across seven hospitals), and Cedars-Sinai all announced enterprise-wide deployments embedded directly inside Epic.
UpToDate, Doximity and a long tail of geography-specific competitors are all racing into the same category from different starting points.
Wave 3 is also in the early majority, advancing rapidly. The thing worth noticing is where it is being adopted: it is moving from individual physicians on their own phones to enterprise deployments embedded in the EHR, at the patient level, with the full context of the patient chart.
Context is key…while patient specific context is vital, “healthcare system specific context” is just as vital. This is significant and it is the bridge to Wave 4.
Here is where most analyses of clinical AI currently stops. They identify three waves, declare external knowledge as the natural endpoint, and move on.
This misses the most important wave of all.
External medical evidence, the kind OpenEvidence and UpToDate vitally act to serve, is what medicine knows. It is the global, published, consensus body of clinical evidence. It is essential. It is also, by definition, the same in every hospital in the world and that is not how clinical decisions actually get made.
Globally, there are approximately 2 billion medical decisions made every day by healthcare professionals (give or take), none of these decisions are made in a vacuum. They are made within complex healthcare systems that have different ways of working, different referral pathways, different formularies differences, differences in access to diagnostic, imaging and staffing facilities etc etc.
This is internal knowledge. And right now, in every health system, in every country, it is either locked away, inaccessible in software not built for the age of AI, in SharePoint sites and shared drives that clinicians don't open or simply stored as “tacit knowledge” in the heads of more experienced team members.
Unlocking this is the next wave.
To use a non healthcare example to put this into scale… JP Morgan, one single company, has over 150 petabytes of internal knowledge and data…GPT 5 was trained on all the data OpenAI could get their hands on, this amounted to 0.28 petabytes.
One single company has over 500x the amount of internal knowledge than one of the most “intelligent AI models ever built”.
Models, trained on external knowledge represent the tip of the tip of the iceberg of knowledge that needs to be unlocked, internal knowledge is the next wave.
Internal knowledge is structurally harder than the three waves before it. That is precisely why it has taken longer to arrive.
The world's medical literature is curated, reasonably structured, indexed, peer-reviewed, and growing predictably. Health systems’ internal knowledge is messy, fragmented across PDFs and intranets, dynamically changing, governed by committees that meet quarterly, and written by clinicians who have day jobs, as well as this, so much of this knowledge lives in clinicians heads.
The "content" doesn't sit in a single source, it sits in 600 different places. There is no editorial board. There is no standard taxonomy. There is no API.
That is exactly why it builds a moat. Nobody outside your health system can replicate your internal knowledge. Once a platform has ingested it, structured it, governed it, and turned it into a trusted AI answer engine that clinicians actually use, the data layer underneath becomes one of the most defensible assets in clinical AI.
We are seeing the first signals of Wave 4 starting in earnest. OpenEvidence's enterprise Epic deployments at Sutter, Mount Sinai, and Cedars-Sinai now bring patient-level context, what the chart says, into the answer engine.
That solves one layer of context.
But there is a second, larger layer of context: health-system-level context, i.e. Internal Knowledge. What does this hospital do, with this kind of patient, on this kind of presentation, i.e. what formulary medications, diagnostic options and referral pathways are available in this health system for this exact patient.
That layer cannot be answered by an external knowledge engine. It has to be answered by a platform that has ingested, structured, and activated the health system’s own internal knowledge.
If you map all four waves onto Everett Rogers' diffusion of innovations curve, and overlay Geoffrey Moore's chasm between early adopters and early majority, the picture looks like this:
Wave 1 (imaging) is in the late majority. Wave 2 (ambient AI) and Wave 3 (external knowledge) are in the early majority, having crossed the chasm. Wave 4 (internal knowledge management) is in the innovators, only now beginning to be adopted by the most forward-leaning health systems.
The arrow on the diagram is the only thing that matters. It is the direction of travel.
Three implications matter for any CMIO, CMO, CNIO, CNO, or Chief Pharmacist:
For twenty years, every health system has tried to solve internal knowledge with the tools it already had i.e. SharePoint, intranets, shared drives, custom portals, occasionally a dedicated policy management system. None of them were built for the way clinical care actually happens, and all of them suffer from the same fatal symptom: clinicians don't use them. A guideline no one reads cannot reduce variation. A protocol no one finds cannot improve outcomes. An audit trail no one populates cannot defend you in court.
For most of the last twenty years, this was a tolerable problem. It is no longer tolerable in 2026 because the rest of the AI stack has moved forward, and internal knowledge is now becoming the bottleneck.
It is tempting to assume that the AI answer engine that wins external knowledge will also win internal knowledge, this is not necessarily true. They are different products, built on different content, governed by different bodies, used in different moments of care. The right architecture is for the internal knowledge platform to integrate with, not be replaced by, the external knowledge platforms your clinicians already use. That is exactly the integration model the most thoughtful health systems are now starting to ask for.
Whichever platform succeeds in being the institutional layer for internal clinical knowledge will accumulate something the other waves cannot: a structured, governed, continuously updated representation of how each health system actually practises medicine. That dataset will be the most strategically valuable asset in clinical AI, and it will be built one health system at a time, beginning now.
The health systems that move first on Wave 4 will define the standards, set the integration patterns, and influence the partnerships. The ones that wait will inherit them.
We've spent the last 4 years quietly building the platform for Wave 4. Today, Eolas is the internal knowledge management platform used by over 400,000 clinicians across 500 sites including approximately 90% of acute NHS trusts, every health board and trust in Wales and Northern Ireland, leading Irish health systems, and US health systems including Stanford, Mass General, Boston Children's, University of Kentucky, and Mercy.
We are the platform that turns your hospital's protocols, guidelines, and pathways into a trusted AI answer engine clinicians actually use, improving the quality and consistency of care.
%201.svg)
%201.svg)
