Is an AI native electronic health record possible?

Nelson Advisors
9 minutes ago
9 min read

Purpose built AI electronic medical record ingesting and analysing multiple data points like wearable technology and remote monitoring in real time

For most of the last decade this was a thought experiment. In 2026 it is a live commercial question and the answer investors need is not “yes eventually” but something more useful: what is technically achievable now, what is structurally blocked and where the value actually accrues.

The short version: an AI native EHR is buildable and parts of it are already shipping. But the constraint has stopped being model capability. It is now data plumbing, regulatory conformity, liability and most decisively, the fact that nobody has solved how to get paid for continuous intelligence rather than discrete encounters.

What "AI native" actually means

The phrase is doing a lot of work in vendor marketing and it is worth separating three quite different claims.

The weakest version is AI-enabled: a conventional EHR with models bolted onto the edges. Ambient scribing, inbox draft replies, coding suggestions. This is where the overwhelming majority of deployed value sits today and it is genuinely useful, but the underlying data model is unchanged.

The middle version is AI-first interface: the record still works the way it always did, but the primary way clinicians interact with it is conversational or agentic rather than click-driven. Oracle Health’s rebuild, constructed on Oracle Cloud Infrastructure rather than extended from legacy Cerner code, with voice-led navigation and embedded clinical agents, is the clearest incumbent example, with the ambulatory product live in the US and acute care functionality scheduled for 2026. athenahealth has taken a similar direction, pairing an ambient scribe with a clinical copilot sitting on an intelligence layer that pulls from other EHRs, payers and registries, with testing running through the first half of 2026.

The strongest version, the one the question is really asking about, is AI native at the data layer: a record whose primitive unit is not the signed clinical note but the continuously updated patient state, assembled from every available signal, with the note generated as a derived artefact when someone needs one. That is a fundamentally different architecture, and essentially nobody has shipped it at scale.

Why the data layer is the hard part

Conventional EHRs are transactional systems designed around billing events. The encounter is the atom. Everything is timestamped to a visit, structured for a claim and optimised for retrieval by a human reading a chart.

A record designed for continuous multi-modal ingestion needs almost the opposite properties. It needs to be event-streamed rather than document-based. It needs a temporal data model that can represent “this patient’s resting heart rate has drifted up 8bpm over three weeks” as a first-class object rather than as something a clinician has to infer by scrolling. It needs provenance on every data point, device, firmware, calibration state, confidence, because a blood pressure from a validated cuff and one from an optical wrist sensor are not the same evidence. And it needs an audit trail granular enough to reconstruct why an agent surfaced what it surfaced, eighteen months later, in a courtroom.

None of that is exotic engineering. Financial services solved comparable problems years ago, and telemetry-heavy industries have run event-streamed architectures at far greater volume for longer. What makes it hard in healthcare is that the data has to come from somewhere, and it mostly lives inside systems with no commercial incentive to release it cleanly.

There is also a subtler design problem. A record built around continuous state has to decide what it believes. When a wearable-derived respiratory rate contradicts a nurse-recorded observation, something has to reconcile them, and that reconciliation is a clinical judgement encoded in software. Conventional EHRs dodge this by storing everything and letting the clinician arbitrate. An AI native system that surfaces a synthesised patient state has taken on that arbitration, along with the accountability that comes with it.

Interoperability has improved, FHIR is now genuinely usable, TEFCA has moved from framework to functioning exchange, and information blocking enforcement has teeth it did not have three years ago.

But the practical experience of assembling a longitudinal patient record from multiple sources is still closer to forensic reconstruction than to querying a database. Data gravity favours the incumbent and Epic’s position, north of 40% of US acute care beds and rising, with roughly 85% of its customer base using at least some of its AI suite, means the most complete datasets sit inside the system with the least reason to make them portable.

The wearable and remote monitoring problem

This is where analytical honesty matters most, because the pitch decks are considerably more confident than the evidence.

The appeal is obvious. A patient generates thousands of data points a day; a clinician sees them for fifteen minutes every six months. Closing that gap with continuous signal and automated interpretation should be transformative for chronic disease management. And there is now a reimbursement pathway: the 2026 CMS changes materially lowered the barriers, cutting the device monitoring threshold from sixteen days to as few as two, introducing a mid-range device code, and creating a time-based code that pays from ten minutes of clinical service rather than twenty. That is a meaningful signal about direction of travel.

But three problems remain unresolved, and they are not model problems.

Signal quality in low-prevalence populations

Consumer wearable atrial fibrillation detection is the best-studied case, and it is instructive. Headline accuracy figures look excellent, the Fitbit Heart Study reported a positive predictive value above 98%, but that figure depends on reflex confirmation with medical-grade ECG. Screening asymptomatic populations with low pretest probability drives false positives up sharply, and specificity for arrhythmias with regular R-R intervals is poor. Bayes does not care how good your model is. Push any detector into a population where the condition is rare and the majority of your alerts will be wrong.

Nobody knows what to do with the output

There are still no guideline recommendations covering how clinicians should act on consumer-grade device data. That is not a gap an EHR vendor can close with better UI. It is a clinical governance vacuum, and it transfers risk directly onto whoever built the system that surfaced the alert.

Continuous monitoring creates continuous duty

This is the constraint most often underpriced in investor conversations. The moment a system ingests a patient’s data in real time, a reasonable person can ask what happened between the signal appearing and someone acting on it. Batch-and-review at least has defensible boundaries. “Real time” invites the question of why the deterioration flagged at 3am was reviewed at 9am. The literature already warns that consumer device volume risks overwhelming an unusually strained clinical workforce; an architecture that ingests everything without a corresponding triage and escalation model does not reduce burden, it relocates and amplifies it.

Any credible AI native EHR therefore needs an opinionated filtering layer as a core competency, not a feature, something that decides what constitutes a clinically actionable change in patient state and, crucially, defends that threshold to a regulator. That is a harder engineering and clinical problem than the ingestion itself, and it is where the genuine defensibility lies.

Four structural constraints

Regulation

This is the immediate one. Full high-risk obligations under the EU AI Act, conformity assessment, technical documentation, post market monitoring, incident reporting, become enforceable in August 2026. Notified body capacity is a real bottleneck and clinical decision support that goes beyond information display sits squarely in software as a medical device territory in the US and UK too.

The commercial implication is unglamorous but important: continuous model improvement collides with a regulatory regime built around versioned, assessed, frozen artefacts. Vendors who architected for weekly model updates are discovering the cost of change control.

Reimbursement

Healthcare buys episodes. An AI native record’s value proposition is continuous, earlier detection, avoided admissions, better population management. In fee for service that value largely accrues to the payer while the cost sits with the provider. The addressable market for a genuinely continuous record is therefore not “all EHR spend” but the much smaller slice operating under real capitated or risk bearing arrangements. That is where the early buyers are, and it is a considerably narrower funnel than the category-level TAM slides suggest.

Switching costs

EHR replacement is a multi-year, eight to nine figure exercise that consumes a health system’s entire change capacity. Being 30% better is irrelevant against that. This is why the greenfield opportunity is concentrated where switching costs are low or the incumbent is weak: new care models, hospital-at-home, specialty and behavioural health, virtual-first primary care, and markets outside the US where the installed base is fragmented.

Liability and evidence

The trial evidence base for ambient documentation is now reasonably encouraging on time saved and burnout. The evidence base for autonomous or semi-autonomous clinical action derived from continuous multi-modal data is thin. Prospective outcome data, not retrospective accuracy metrics, is what will unlock enterprise procurement, and generating it takes years.

A note on the UK and European picture

The structural constraints look different outside the US, and in ways that cut both directions. The absence of a fee-for-service claims layer removes the reimbursement obstacle almost entirely, an NHS trust or an integrated European payer-provider captures the value of an avoided admission directly, which is precisely the alignment a continuous record needs. Single payer systems are, in principle, the natural first buyers.

Working against that is procurement velocity, capital constraint and a fragmented supplier base with deep incumbency in individual trusts. And from August 2026 the EU AI Act adds a compliance burden that US-domiciled competitors can defer.

The realistic European play is therefore not a general-purpose AI native EHR but a condition specific or pathway specific one, heart failure, COPD, diabetes, frailty, where the monitoring signal is well characterised, the clinical guidelines already exist, and the avoided-cost case can be evidenced inside a single budget holder. That is a smaller initial market with a considerably shorter route to proof.

So, is it possible?

Yes, but it will not arrive as a rip & replace and investors betting on an “Epic killer” are probably mis pricing the shape of the outcome.

The likelier path is a decoupling. The legacy EHR persists as the system of record, the regulated, certified, billing-integrated substrate nobody rips out. Meanwhile an AI layer becomes the system of engagement, absorbing the clinician’s actual working surface and progressively the system of intelligence, holding the continuously updated patient state that the underlying record cannot represent.

The economics of that layer are what matter. Abridge is the cleanest proof point: roughly $100m ARR, a $5.3bn valuation as of mid-2025, a further $316m raised in April 2026, and deployment across ninety-plus disclosed health systems including Kaiser and Mayo. That was built on a single workflow. The company that generalises from documentation to continuous patient state has a substantially larger prize.

Two things determine whether that layer becomes a durable business or a feature. First, whether the incumbents can commoditise it, Epic’s Agent Factory and its Curiosity foundation models, trained on anonymised real world records, are an explicit attempt to do exactly that and distribution to 85% of an installed base is a formidable weapon. Second, whether continuous monitoring produces outcome evidence strong enough to change reimbursement. If it does, the pull comes from payers rather than providers, and the buying centre shifts entirely.

For anyone allocating capital in this space, the diligence questions worth pressing are narrow and specific. Does the product own a proprietary data asset or merely a workflow on someone else’s? What is the regulatory classification and has conformity assessment actually started?

Where is the clinically validated triage logic that stops continuous ingestion becoming an alert firehose? Is there a defined escalation pathway with named clinical accountability? And is the buyer bearing risk, or paying fee for service?

The technology is no longer the binding constraint. The question is whether the system around it can be reorganised fast enough to pay for what the technology can already do.

Nelson Advisors > European HealthTech, MedTech, Digital Health Investment Banking

Nelson Advisors specialise in Mergers and Acquisitions, Partnerships and Investments for Digital Health, HealthTech, MedTech, Health IT, Consumer HealthTech, Healthcare Cybersecurity, Healthcare AI companies. www.nelsonadvisors.co.uk

Follow Nelson Advisors LinkedIn Page > https://www.linkedin.com/company/nelson-advisors/

Nelson Advisors regularly publish Thought Leadership articles covering market insights, industry trends, deal commentary, market analysis & predictions. https://www.healthcare.digital

Nelson Advisors publish Europe's Leading Healthcare Technology Investment Banking Newsletter every week, join 5000+ HealthTech and MedTech subscribers today! https://lnkd.in/e5hTp_xb

#VentureCapital #PrivateEquity #Founders #SeriesA #SeriesB #Founders #SellSide #TechAssets #Fundraising #BuildBuyPartner #GoToMarket #PharmaTech #BioTech #Genomics #NelsonAdvisors #HealthTech #MedTech #DigitalHealth #HealthIT #Cybersecurity #HealthcareAI #FemTech #Mergers #Acquisitions #Partnerships #Growth #Strategy #NHS #UK #Europe #USA #Canada