Ambient AI's specialty expansion: oncology, surgery, cardiology, and the TAM question for 2027

Ambient AI documentation won its first category by solving the simplest version of the problem. Primary care, hospital medicine, and parts of urgent care generate large volumes of relatively standardized clinical encounters, dominated by history-taking and physical examination, with notes that follow a recognizable SOAP structure and little specialty-specific terminology that a general-purpose language model cannot handle. The two leading vendors — Abridge in the independent-system market and the Nuance/DAX stack in the Microsoft-oriented market — reached scale on this foundation, and by mid-2026 ambient AI in primary care and hospital medicine had moved from early-adopter curiosity to default workflow at most of the major U.S. academic medical centers and large community systems.

The specialty expansion that both vendors are now pursuing is a different product problem, and the vendors that execute it well will substantially expand their addressable market. Those that execute it poorly will find that the specialty market resists the general-purpose ambient model in ways that the primary-care market did not.

Three specialties, three product problems

The three specialties where ambient AI expansion is most actively being pursued — oncology, surgery, and cardiology — share the characteristic that their note structures are substantially more specialized than primary care notes. But the specific product problem each presents is different.

Oncology notes are dominated by structured data that is not natively captured in the audio of an encounter. The oncologist’s clinical reasoning in a tumor-board or treatment-planning encounter references staging data, biomarker results, imaging interpretations, prior treatment regimens with specific doses and dates, response assessments, and toxicity grades. Much of this is in the EHR but not spoken; much of it is spoken in abbreviated shorthand that an ambient system has to resolve against structured sources to produce an accurate note. The product problem is less “transcribe and summarize the conversation” and more “fuse the conversation with structured oncology data to produce a note that reads correctly to another oncologist.” Vendors that solve this well will produce notes that oncologists accept with minimal editing. Vendors that only transcribe and summarize will produce notes that oncologists reject as missing the clinical context that makes the note useful.

Surgical notes have a different structure entirely. The operative report is the authoritative document, and it is structured around anatomy, surgical approach, steps performed, findings, complications, and specimens. The patient encounter that an ambient system might capture (the pre-op visit, the post-op visit, the operative report dictation) is not the same across phases, and the ambient tooling has to understand which phase of the surgical workflow it is operating in. For intra-operative dictation, the system needs to handle a different acoustic environment (operating room noise, team conversation, equipment sounds) and a different vocabulary. For the pre-op and post-op visits, the system can look more like a primary-care ambient tool. Vendors expanding into surgery will likely need separate product configurations for the different workflow phases, which is operationally more complex than a single primary-care product.

Cardiology notes are the most measurement-dense of the three. A cardiology note typically references specific ejection fraction values, specific chamber sizes, specific valve gradients, specific device interrogation parameters, specific stress-test results, and specific hemodynamic calculations. Much of this is quantitative and much of it is captured in device output rather than in the clinical conversation. The product problem is extracting and fusing device output (echo reports, cath lab reports, device interrogations, stress-test outputs) with the clinical conversation to produce a note that reads with the measurement precision cardiologists expect. The specialty is also internally stratified — general cardiology, interventional, electrophysiology, heart failure — with different sub-specialty vocabulary and note structures.

Why general-purpose ambient AI struggles in specialties

The core reason a general-purpose ambient AI product that excels in primary care may underperform in a specialty is that the primary-care note is almost entirely a reconstruction of the clinical conversation, while the specialty note is a reconstruction of the clinical conversation fused with specialty-specific structured and semi-structured data from the EHR and from upstream clinical systems. The ambient system that captures only the audio is working with a fraction of the information that produces the final note.

The engineering fix is a specialty-specific data fusion layer that pulls from oncology-specific sources (tumor registries, pathology, molecular diagnostics, prior regimens), surgical-specific sources (OR scheduling, pre-op workup, intra-operative monitoring), or cardiology-specific sources (echo lab, cath lab, device interrogation systems, stress testing). Building this fusion layer is 12-18 months of engineering per specialty, and it has to be built against the specific EHR customer base — the Epic implementation of oncology data is different from the Oracle Health implementation, which is different from the Meditech implementation.

The incumbent specialty-focused documentation vendors are not sitting still. Oncology-specific documentation tools from Flatiron, Tempus, and others have been iterating for years and have the specialty data fusion muscle that the general-purpose ambient vendors are now trying to build. The question for the general-purpose vendors is whether they can build the fusion layer faster than the specialty incumbents can add ambient-AI capability.

The TAM mathematics

The specialty expansion matters commercially because the TAM for ambient AI in primary care, hospital medicine, and the other “general” workflows is substantial but finite. The U.S. has roughly 230,000 primary care physicians, roughly 70,000 hospitalists, and some additional capacity in urgent care, emergency medicine, and the ambulatory-specialty “generalist” segment. Ambient AI penetration into this market segment is now high enough that the per-vendor growth curve is bending — the next 100% of market share is substantially harder to acquire than the first 100%.

The specialty market is additive in a meaningful way. Oncology adds roughly 15,000 medical, surgical, and radiation oncologists in the U.S. Surgery adds roughly 230,000 surgeons across general and specialty categories. Cardiology adds roughly 35,000 cardiologists across sub-specialties. Each specialty, if captured, is tens of billions in incremental ARR potential across the vendor landscape, and each specialty that is not captured by the general-purpose vendors remains available to specialty-focused competitors.

The strategic calculation for the leading vendors is which specialties to prioritize. Oncology has the strongest documentation-pain and economics narrative — the per-encounter note is complex, the time cost on the oncologist is high, and the specialty’s economics support higher per-seat pricing. Surgery has the largest absolute physician count but the most workflow-complexity to crack. Cardiology has the highest device-fusion complexity but the strongest EHR-integration leverage because cardiology data is already well-structured in the leading EHRs.

Deployment patterns to watch

Three deployment patterns in Q4 2026 will signal which vendors are gaining specialty traction.

The first signal is specialty-specific product branding. Vendors that are serious about a specialty will give the specialty product its own name, its own product manager, its own clinical advisory council, and its own release cadence. Abridge, DAX Copilot, Ambience, Nabla, and others are each at different stages of this specialty productization, and the trajectory over the next two quarters is going to tell the market who is serious.

The second signal is specialty-society partnership. Specialty-society endorsements — ASCO for oncology, ACC for cardiology, ACS for surgery — are not required for deployment but are required for the specialty’s center-of-gravity institutions to feel confident adopting. The vendor that signs the first ASCO or ACC endorsement for its specialty ambient product will have substantial momentum.

The third signal is specialty-specific outcome data. Primary-care ambient AI deployments are now being published with real after-hours documentation reduction, note-completion-time reduction, and clinician-satisfaction data. The same evidence generation needs to happen in each specialty. The vendor that publishes the first peer-reviewed oncology or cardiology deployment outcome paper will set the evaluation standard for the category.

Outlook

By the end of 2027, ambient AI in specialties will no longer be an aspiration — it will be a differentiating capability or a competitive liability. The vendors that have built the specialty data fusion layer, signed specialty-society partnerships, and published specialty deployment outcomes will be in a materially stronger market position than those that are still extending the primary-care product into specialty use cases. The TAM expansion that successful specialty execution unlocks is in the tens of billions. The market concentration that unsuccessful specialty execution produces will reward the vendors that got there first.