How recursive reasoning could enhance patient safety
By Lars Maaløe, Co-Founder and CTO, Corti
Healthcare AI is at a tipping point. Ambient scribes and virtual assistants were meant to lighten documentation burdens, giving clinicians more time for patients. Instead, new layers of administrative complexity have emerged, leaving clinicians to correct AI outputs, manage new risks, and navigate unfamiliar workflows while remaining ultimately liable for the record.
The limitations are as much conceptual as technical. Current systems treat clinical conversations as static blocks of text to be transcribed after encounters end. But clinical encounters are dynamic, nonlinear, and riddled with ambiguity. The patient’s story unfolds over time, facts emerge in pieces, and context can shift instantly.
From summarisation to clinical reasoning
What’s missing isn’t better transcription – it’s clinical thinking. Physicians don’t merely record what was said, they reason through it. They interpret symptoms, weigh probabilities, resolve contradictions, and construct meaning collaboratively with patients.
Recursive reasoning offers a compelling alternative. Instead of processing completed conversations, these systems engage in real time, identifying and refining clinical facts as they arise. Each fact is validated against multiple layers of context – what’s been said, the patient’s history, known clinical pathways – before being structured into the note.
This approach mirrors how clinicians work: gathering signals, making sense of inconsistencies, and continuously adjusting understanding. The result reflects actual clinical reasoning and gives clinicians confidence.
Transparency over trust
One central challenge in healthcare AI is the trust gap. Clinicians are rightfully cautious about systems delivering polished outputs without showing their work. Recursive systems address this by making their logic visible. Clinicians can see the reasoning process unfold, intervene where needed, and course-correct during encounters rather than after.
This transparency reframes accountability. Rather than placing validation burden entirely on clinicians at the visit’s end, it distributes oversight throughout the process. This aligns more closely with clinical standards of care and emerging regulatory frameworks demanding explainability and auditability.
Beyond the note
The broader implications are intriguing. The same infrastructure supporting real-time documentation can power other clinical tools – from differential diagnosis aids to context-aware alerts. This represents more than note-taking evolution; it points to a new model for AI participation in clinical workflows.
Imagine an AI that doesn’t just transcribe chest pain mentions, but simultaneously flags missed ECGs, checks medication interactions, and reminds clinicians of prior abnormal troponin levels – all within natural conversation flow. That’s the potential when reasoning, not transcription, is the foundation.
Implementation with eyes wide open
Real-time systems require specialised infrastructure and place greater demands on integration, latency management, and validation protocols. But these challenges aren’t insurmountable or unique to recursive approaches.
What sets this path apart is that added complexity directly serves the core objective: safer, more reliable clinical care. Unlike general-purpose AI adapted to healthcare’s edge cases, recursive systems are designed from the ground up with clinical reasoning in mind.
Success will depend on thoughtful deployment. Systems must complement existing workflows, be transparent and editable, and respect clinicians’ central role as interpreters of meaning, not just correctors of mistakes.
The road ahead
The promise of AI in healthcare isn’t about automation alone—it’s about building tools that genuinely support the cognitive work of care, organizing and clarifying the flood of information clinicians face daily.
Recursive reasoning isn’t a silver bullet, but it may represent the beginning of a new phase where AI systems think alongside clinicians, not merely document after them.
In doing so, they have potential to restore what the first wave of healthcare AI often overlooked: trust, transparency, and alignment with the realities of clinical practice.
