From Ambient AI to Clinician-Directed Intelligence

By Matt Troup, PA-C and Chaitanya Asawa

AI has made it possible to draft documentation with impressive speed. In many cases, it reduces administrative effort. Yet speed alone does not resolve the central issue: clinicians need the final output to reflect their reasoning and priorities. The gap is no longer “Can AI generate a note?” but rather, “Does the note communicate what the clinician intended?” That distinction is clinical, but it is also deeply technical. From a software engineering standpoint, it shifts the problem from producing fluent text to producing output that is contextual, steerable, and stable under refinement.

At Abridge, we have been working toward a specific balance: ambient automation that feels effortless, paired with clinician control that feels natural. Documentation should not require undue burden on the clinician, and the draft should not require reconstruction after the fact. But natural automation at this depth can only succeed with controls. Agency is what keeps that automation clinically valid. The note ultimately belongs to the clinician, and the system has to make it easy for the clinician to shape what ends up in the chart.

When that balance is right, clinicians stop feeling like they’re using a tool to serve workflows and begin to feel more like conductors guiding an orchestra of capabilities. The system can handle the background complexity: structuring the note, adapting to visit type, carrying forward context, all while the clinician remains responsible for emphasis, interpretation, and judgment. Importantly, this is not an argument for more work. It is an argument for different work: fewer mechanical edits, more intentional direction. When the interaction is designed correctly, the time spent “fixing the note” is replaced by time spent aligning the note with clinical intent.

This is the philosophical foundation for what we are launching today: The ability to edit the note using prompts written in natural language. Prompt editing gives clinicians a simple way to direct the note, without turning the finalization process into manual rewriting. In day-to-day use, the prompts feel familiar because they are grounded in the same mental checklist clinicians already run as they review and refine their documentation.

A clinician might say, “Make the assessment more concise and make the differential explicit,” because that is how they think and how they want their reasoning represented. Another might say, “Clarify why we chose outpatient follow-up rather than imaging today,” because the documentation needs to carry that decision-making forward. Another might say, “Add 'Next Steps' section below the A/P that clearly calls out the patient's action items,” making the treatment plans easier for colleagues to understand at a glance. The clinician is no longer confined to accepting a draft and patching it manually. The clinician can direct and customize the output.

Prompt editing, however, is not the end of this story. It is only the first step in a broader shift toward clinician-directed intelligence in the flow of care, and this is where Clinical Decision Support becomes central. Decision support is not new; clinicians have used evidence-based resources for years, but there is a barrier. Decision support often lives outside both the clinical workflow and the documentation workflow. It may provide the right information, but it forces clinicians to leave the environment where they are thinking, determine what is relevant, then translate it back into the note in a way that is defensible and clear. That translation step is exactly where both friction and risk accumulate.

Our thesis is that the next generation of decision support should be embedded, contextual, transparent, and documentable. Embedded means it lives where the clinician is already working, rather than in a separate tab. Contextual means it is grounded in the patient and the encounter, rather than requiring the clinician to restate details that are already present in the conversation or chart. Transparent means the evidence is inspectable, with clear provenance and citations. Documentable means that when evidence changes the plan, that evidence can be efficiently incorporated into the final note in the clinician’s voice, rather than requiring copy/paste and rework.

This is why we have invested so heavily in building decision support around trusted clinical content, particularly by contextualizing evidence-based insights with the conversation transcript and Abridge-generated note. The promise here is not “ask a question and receive an answer.” The promise is that a clinician can ask a question in the middle of note review and receive an evidence-grounded response with citations. Then, the clinician can decide how it should or should not influence the plan and the documentation. This sequence matters. It keeps the clinician in control, and it keeps evidence from becoming a separate activity that is divorced from the record.

Prompt editing is the connective tissue that makes this practical. When decision support is paired with agency, the clinician can move from evidence to documentation without friction. A clinician may review a contextualized response from UpToDate, decide that it changes how they want to communicate a plan, and then use a prompt to integrate that reasoning in a way that is accurate and stylistically consistent. This is fundamentally different from bolting a reference tool onto the side of a note. It is decision support that respects judgment and actually strengthens documentation.

From the AI side, the same principle applies: clinical decision support in this setting must be built for trust and clinician control. The clinical decision support agent retrieves information from trusted content sources and grounds its responses with clear, precise references. When sourcing this information, the agent leverages the encounter to contextualize its search results within the patient’s circumstances.

Trust and clinical integrity are central to how clinical decision support is designed and evaluated. Abridge conducts internal evaluations across clinical quality and clinical safety to measure aspects ranging from precision, recall, and faithfulness to source material as part of its product development and quality processes. 

At Abridge, prompt editing is our first clear expression of agency in ambient documentation. Embedded decision support, grounded in trusted content and contextualized to the patient and encounter, is the next major step. Over time, this approach opens the door to a workflow that feels more coherent: documentation that is effortless when it should be, and intentionally shaped when it must be. This is a future where clinicians are not asked to choose between speed and authorship.