AI Voice Agents in the Cath Lab: What the Sofiya Pilot Means for Practice
A 90-day pilot of an agentic voice assistant handling pre-procedural calls offers a data point on where administrative AI is heading in interventional cardiology — and which public companies sit near the trend.
A high-volume academic cath lab has spent the past year testing whether a voice-based AI assistant can safely take over one of the most repetitive tasks in pre-procedural care.
The assistant, an agentic conversational system built specifically for the workflow, calls patients the day before catheterization to review logistics, confirm allergies and medications, and answer routine questions.
Two sequential 90-day phases at the Mount Sinai Fuster Heart Hospital cath lab generated the largest published experience with this kind of tool in interventional cardiology to date.
For a specialty facing a worsening nursing shortage, the results are worth a close read.
Why This Matters Now
Preprocedural calls are clinically essential but highly repetitive, covering arrival time, fasting instructions, transportation, and a structured allergy and medication review.
In a lab performing more than 16,000 procedures a year, that workload consumes several full-time nursing hours every single day.
Workforce shortages projected to worsen through 2030 have pushed health systems to look for ways to protect nurses' time for direct, bedside clinical work.
The pilot's central question was whether a large language model-based voice agent could take on this task without compromising safety or the patient experience.
What the Pilot Found
Across two 90-day phases, 1,431 patients received 1,606 calls, and the overall successful-completion rate rose from 86.4% in phase one to 87.9% in phase two.
More than a third of all calls were handled entirely by the AI agent with no nurse involvement at any point.
Roughly another third triggered a brief, protocol-defined nurse callback, and about 12% were escalated to a full human conversation.
Investigators estimated the assistant saved an average of 11 minutes per call, equivalent to more than 37 twelve-hour nursing shifts over a year in this single lab.
Patient satisfaction was high in both phases, and only a small minority of patients declined to speak with the AI system at all.
Six calls were derailed by an AI hallucination — the system generating inaccurate information — out of more than 1,600 attempts, and every transcript was reviewed by a nurse before anything reached the chart.
| Metric | Phase 1 (n=806 calls) | Phase 2 (n=800 calls) |
|---|---|---|
| Successfully completed calls | 86.4% | 87.9% |
| Fully automated, no nurse contact | 36.6% | 42.6% |
| Brief protocol-driven nurse callback | 37.6% | 33.5% |
| Escalated to full human call | 12.2% | 11.8% |
| Patient declined AI call | 4.3% | 2.8% |
| Postprocedural satisfaction | 94.7% | 98.1% |
Case Vignette
A 71-year-old woman with stable angina is scheduled for diagnostic catheterization the next morning.
The evening before, an AI voice agent calls to confirm her arrival time, review same-day fasting rules, and ask about allergies and current medications.
She mentions she takes a direct oral anticoagulant that was not listed in her chart, and the system automatically flags the discrepancy for a same-day nurse callback rather than attempting to resolve it itself.
A nurse calls back within the hour, confirms the hold parameters with the proceduralist, and documents the correction — illustrating the intended division of labor between automation and clinical judgment.
Where the Guardrails Are
The design keeps a human in the loop by default: every transcript is reviewed by nursing staff, and any question outside the system's scripted knowledge base triggers a callback rather than an improvised answer.
At the start and end of every call, the assistant identifies itself as a virtual agent and reminds patients that a person is always reachable.
Deployment reportedly required review by roughly a dozen institutional committees, including the health system's first case brought before a dedicated AI governance board.
That governance overhead is worth noting for any practice considering a similar build, since the technical lift is often smaller than the compliance one.
The Investor Angle
The specific platform used in this pilot is an internally customized, vendor-partnered build rather than a standalone public product, so there is no direct ticker to attach to it.
The broader trend it represents, though, touches several publicly traded companies that physician-investors may already be watching.
| Company | Ticker | Relevance | Analyst Consensus | 12-Mo. Price Target |
|---|---|---|---|---|
| Sofiya voice-agent platform | no ticker (private) | Custom-built, vendor-partnered agentic AI; not a standalone public offering | — | — |
| Microsoft | NASDAQ: MSFT | Ambient and agentic clinical AI (Dragon Copilot) competes in the same workflow-automation category | Strong Buy | $558.77 (+39.3%) |
| Tempus AI | NASDAQ: TEM | Pure-play healthcare AI/data company scaling clinical workflow and diagnostics products | Buy | $66.06 (+12.6%) |
| GE HealthCare | NASDAQ: GEHC | Cath lab imaging and interventional-suite equipment maker layering AI into existing hardware | Buy | $79.33 (+23.1%) |
None of these companies were named as the vendor behind this specific pilot, and the table above should be read as thematic exposure to administrative and clinical AI in cardiology, not as a claim about who built Sofiya.
What Might Come Next
Program leaders have described plans to extend similar AI calling to postprocedural follow-up, appointment scheduling, and cardiac rehabilitation referrals.
Expansion into peripheral and structural heart disease scheduling within the same cath lab has also been discussed as a next step.
Nursing representatives have publicly cautioned that every AI-generated summary still needs a clinician's review before it becomes part of the permanent record.
That tension between efficiency gains and continued oversight is likely to define how quickly this category scales across other cath labs.
Bottom Line
A well-governed, human-in-the-loop AI voice agent handled more than a third of pre-procedural cath lab calls with no nurse involvement and freed an estimated 37-plus nursing shifts a year at a single high-volume site.
The technology performed safely within a narrow, scripted scope, but every escalation pathway and every hallucination in this pilot was caught because a nurse was still reviewing the output.
For practices evaluating similar tools, the governance and review infrastructure may matter as much as the underlying model.
Further Viewing
▶ Harnessing Artificial Intelligence at the Mount Sinai Health System
▶ Mount Sinai Department of Artificial Intelligence (AI) and Human Health
References
- Meet Sofiya, Mount Sinai's latest AI assistant. Becker's Cardiology.
- Utilizing an AI-assisted virtual agent for pre-procedural patient calling in the cardiac catheterization laboratory. European Heart Journal – Digital Health, 2026.
- Mount Sinai Launches Cardiac Catheterization Artificial Intelligence Research Lab. Mount Sinai Newsroom.
- Microsoft Corporation (MSFT) stock overview. StockAnalysis.com.
- Tempus AI (TEM) stock forecast. StockAnalysis.com.
- GE HealthCare Technologies (GEHC) stock overview. StockAnalysis.com.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.