The Three-Stage Journey Healthcare Must Take Before Conversational AI Can Safely Enter the Clinic
New analysis reveals why skipping foundational speech recognition steps could put patients at risk as error rates remain stubbornly high in real-world settings.
Healthcare's rush toward conversational AI assistants that can execute clinical tasks on voice command faces a sobering reality check: the technology still struggles with basic accuracy in real-world hospital environments. According to Dan Engel, Principal Product Manager at Corti with 15 years of healthcare AI experience, the industry must master foundational capabilities—accurate speech recognition and intent understanding—before systems can reliably act on clinician requests. With physicians already spending an average of 13 hours weekly on indirect patient care tasks including documentation and order entry, the promise of voice-activated assistance is compelling. But recent error rate data suggests the path forward requires patience and incremental progress rather than technological leaps.
The Accuracy Gap That Won't Close Fast Enough
Despite decades of development in clinical speech recognition, significant accuracy challenges persist that make conversational AI deployment premature in many settings. A 2025 systematic review of 29 studies found AI-based speech recognition word error rates ranging from 8.7% in controlled dictation settings to over 50% in conversational or multi-speaker environments—the exact conditions where conversational AI would operate.
The problem extends beyond background noise. A 2026 study found an 11 percentage point increase in word error rate for non-native English speakers using Whisper speech recognition, highlighting how diversity in clinical teams creates additional accuracy barriers. When systems move from passive transcription to active execution of commands, these error rates become clinically dangerous rather than merely inconvenient.
From Transcription to Action: Why Each Stage Matters
The evolution from back-end transcription to ambient AI scribing to conversational agents represents incremental stages that build upon rather than replace previous technologies, according to Engel's analysis. KLAS survey data shows 79% of U.S. healthcare organizations now use ambient speech for clinical documentation, representing the current frontier of widespread adoption.
Yet even at this intermediate stage, accuracy concerns remain substantial. A 2025 study evaluating two ambient AI scribe products found errors in 70% of AI-generated draft notes, with an average of 2.9 errors per note. These errors occur when systems only need to transcribe and summarize—not execute actions that could directly impact patient care.
Engel's experience developing virtual-assistant intent recognition for physicians in 2017-2018 illustrates the complexity of the next stage. His team expanded from approximately 12 supported use cases to over 100 intents, revealing how the combinatorial explosion of possible physician requests quickly overwhelms system capabilities.
The Integration Challenge No One Wants to Discuss
Beyond speech accuracy and intent understanding lies a third barrier: the gap between AI systems' ability to understand what clinicians want and EHR systems' ability to execute corresponding actions. This integration challenge represents a significant obstacle that limits return on investment and adoption, requiring coordinated development between AI vendors and core clinical systems.
Engel recommends that healthcare IT leaders maintain multiple speech interaction modalities—traditional dictation, ambient scribing, and conversational agents—rather than pursuing single-solution strategies. This approach allows clinicians to select appropriate tools based on task complexity and workflow context, acknowledging that different clinical scenarios demand different accuracy thresholds and interaction models.
The staged evolution of clinical speech technology highlights a critical implementation principle: accuracy thresholds must be substantially higher when systems execute actions versus simply transcribing text, as misunderstood intent can lead to clinical errors. Healthcare organizations investing in conversational AI must resist the temptation to leapfrog directly to voice-activated agents without first establishing robust foundational capabilities.
Sources
Google Search
Follow Health AI Central on Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.