Skip to content
AI in Healthcare

Article

Ambient AI in the OR — surgical video, note capture, and workflow integration

The OR is a different environment for ambient AI than the clinic. Surgical-video AI, intra-op scribing, and where the CPT and workflow shifts are landing.

By AI in Healthcare Editorial Updated
  • ambient-AI
  • surgery
  • surgical-video
  • OR
  • CPT
  • workflow
  • perioperative
  • ambient-scribe

Most of the ambient-AI conversation in 2026 is a clinic conversation. A primary-care physician’s ambient scribe listens to a room-mic-plus-EHR conversation and produces a draft SOAP note. That workflow has become mature enough that nearly every major academic center has piloted one, and multiple enterprise-scale rollouts — the VA’s nationwide 2026 deployment foremost among them — are running against thousands of clinicians in production.

The operating room is a different animal. The data streams are different, the users are different, the documentation obligations are different, and — crucially — the failure modes are different. Ambient AI in the OR has been quietly compounding through 2025 and 2026 into a distinct category that borrows the interface conventions of clinic ambient scribes but rests on a very different technical and clinical foundation. This piece walks through what actually exists in the OR today, where the surgical-video AI market sits, how perioperative ambient scribing is shaping up, and the CPT-coding shifts that will determine whether any of it becomes durably reimbursable.

Why the OR is different

Four things distinguish the OR from the clinic room for ambient-AI purposes:

  • Multiple data streams. A clinic encounter is essentially one audio channel between the clinician and the patient. An OR captures continuous surgical video (from the laparoscopic tower, the endoscope, or an overhead camera), continuous physiologic data from the anesthesia workstation, timeouts and role-checklists from the OR nurse, verbal exchanges between the surgeon, first assist, scrub tech, circulator, and anesthesia team, and (increasingly) audio from a smart-speaker-style OR microphone system. That is at least four distinct input modalities, all clinically important and all producing structured or semi-structured artifacts downstream.
  • Multiple documentation owners. Clinic notes are written by the physician. OR documentation is fragmented across the surgeon (operative note), the OR nurse (intra-op nursing record), the anesthesia provider (anesthesia record), and — for teaching cases — the fellow or resident. Each has its own regulatory and billing expectations.
  • Sterile-field constraints. The surgeon cannot pull out a phone. Any UI has to be voice-driven, foot-pedal-driven, or displayed on a monitor already inside the room. The most successful OR AI deployments assume the surgeon’s hands are on the tools; the AI has to be ambient in a hard sense of the word.
  • Higher stakes for error. A wrong lab value in a clinic note is fixable at the next visit. A wrong instrument count, a missing timeout, or a mischaracterized intraop event has both patient-safety and medico-legal consequences that surface fast.

The surgical-video AI landscape

The oldest thread of OR AI is the surgical-video layer — computer-vision systems that ingest laparoscopic, endoscopic, or robotic video and produce structured outputs downstream. The category has consolidated into a handful of clearly-positioned players:

  • Theator, Caresyntax, and C-SATS (the last acquired by Johnson & Johnson years ago) are the pure-play surgical-video-analytics vendors, each with their own take on real-time phase detection, instrument recognition, event annotation, and post-op debrief.
  • Activ Surgical works on augmented visualization — real-time overlays that surface anatomy the surgeon can’t easily see (perfusion, bile ducts, ureters). This is closer to a device play than an ambient-AI play and it is coming to market through 510(k).
  • Intuitive Surgical, Medtronic Hugo, and other robotic platforms are integrating their own vision-AI stacks into the robot itself, which changes the ownership question for OR AI in robot-heavy service lines.

The evidence base for surgical-video AI took a real step forward with the NEJM AI paper from Saldanha et al. — a decentralized, swarm-learning pipeline that produced patient-level predictions on laparoscopic appendectomy videos on par with centralized training, across six international centers. The paper is doing two things at once: it is a proof point for a specific class of model, and it is a proof point that federated training can reach centralized performance on surgical-video tasks. That second finding matters more, because it undercuts the “we can’t train AI on our surgical video because of privacy” objection that has capped the category to date. Expect the argument to shift from “privacy blocks distributed training” to “we still need governance around who trains what on whose data, even if the pipe is federated.” Which is where the debate belongs.

Ambient scribing in the perioperative envelope

Alongside the video track, the ambient-scribe layer is arriving in the OR — but shaped very differently than it is in clinic. The three deployment patterns that actually work in perioperative settings in 2026:

  • Post-op operative-note drafting. The surgeon dictates or holds a brief post-op debrief; the ambient system produces a structured op note aligned to the specialty template (CPT codes surfaced, ICD-10 diagnoses populated, findings and procedure steps captured). This is the closest analogue to clinic ambient scribes and the deployment pattern most vendors are actually shipping today.
  • Intra-op nursing-record automation. The OR nurse’s intra-op record — timeouts, counts, position changes, medication administration events, timing markers — is a highly structured document that has been captured largely by hand or by point-and-click for decades. Ambient AI that listens to the room and pre-populates the record has cut nurse documentation time meaningfully in early pilots. Vendors here are shipping mostly into Epic OR and Oracle Health OR modules; Epic’s OR module has been adding integration points through 2025–2026.
  • Anesthesia-record adjunct. The anesthesia workstation already emits highly structured data; the ambient layer here is smaller — capturing verbal exchanges (drug names, dose confirmations, response-to-treatment) that don’t get into the workstation stream.

For every one of these deployment patterns, the guardrails around BAA coverage, audio-recording retention, and vendor access to raw audio are more scrutinized than they are for clinic scribes. OR audio captures identifiable non-patient voices (staff conversations, sometimes surgeon-to-team commentary that would be legally sensitive if surfaced) and the governance model has to account for that.

The CPT-coding shift that determines everything

A recurring lesson from healthcare-AI history is that no category matures without a payment model. For OR AI in 2026, the CPT-coding picture is starting to move but is not yet where the vendors need it to be:

  • AMA’s Category III CPT codes for augmented-reality visualization cover several of the imaging-overlay use cases. Category III means tracking codes — they support reimbursement negotiations but do not carry a fixed RVU. Elevation to Category I is the milestone to watch.
  • The 2025–2026 CPT release introduced the first codes explicitly recognizing AI/ML augmentation in imaging interpretation and quantitative analysis. Extension to OR-specific AI-augmented workflows is being discussed by the AMA CPT Editorial Panel, driven in part by comments from the American College of Surgeons.
  • CMS’s own AI-adjustment posture — how the physician fee schedule treats AI-augmented services — matters as much as the code itself. If AI-augmented time is replacing physician time rather than adding to it, the RVU treatment is unlikely to be generous.

Sources cited

Workflow integration: where deployments succeed or fail

In the mature clinic ambient-scribe market, deployment success correlates with EHR integration depth. In the OR, the equivalent lever is device and workflow integration depth. The deployments that succeed do a few specific things:

  • Meet the surgeon at the monitor already in the room. Any output that requires a new screen is friction; outputs that show up on the laparoscopic tower or the anesthesia display fit into how the room already works.
  • Handle the sterile-field constraint gracefully. Voice, foot pedal, or nurse-mediated interfaces. Any hardware in the sterile field has to have real device-clearance and infection-control provenance.
  • Ship a real integration to the OR module. Epic OR, Oracle Health OR, MEDITECH surgical suite, and the anesthesia workstation vendors (GE, Dräger, Mindray). “We export a PDF” is not integration.
  • Sit inside health-system AI governance. OR AI is high-consequence work with heavy medico-legal exposure. Deployments outside the health system’s RUAIH-style governance program will not survive an incident review, and incidents will happen.

What to watch through 2027

  • A first true OR-focused ambient-scribe FDA clearance. Most current OR ambient products live on the “documentation adjunct” side and avoid the device line. If one crosses it — for example, a system that recommends coding levels or surfaces intraop safety alerts — the regulatory conversation gets more interesting fast.
  • Elevation of surgical-AI CPT codes from Category III to Category I. The commercial inflection.
  • A high-profile incident. Category maturity in healthcare AI is usually established by the incident that reshapes the governance conversation. Surgical AI has not had its defining one yet. It will.
  • Robotic-platform vs. third-party vendor equilibrium. Whether Intuitive, Medtronic, and CMR Surgical wall off their video streams or open them to third-party OR AI vendors is a market-defining choice happening quietly through 2026.