Before a visit, clinicians often have to reconstruct a patient’s story from years of notes, results, orders, medication lists, and follow-up encounters. They may have only a minute to do it. An AI summary can make that record shorter, but shorter is not the same as prepared.

The harder problem is deciding what deserves attention now: which facts could change the upcoming decision, how different pieces of evidence fit together, and where the record is still unclear.

At Ambience, we believe good chart preparation should feel like working with a trusted senior resident during a busy clinical day: someone who can review the chart independently, understand why the patient is here, trace how the care plan has changed, and identify what deserves attention today.

That belief is shaped by the people who build the technology. Nearly one in five Ambience employees is an active or trained clinician, with clinical expertise embedded across product development, model training, and clinical validation. Many understand these problems from their own experience in care delivery. They help shape what we build and how we judge whether it works. By taking on more of the investigative work that goes into patient care, we can help clinicians focus their expertise on the decisions and conversations that matter most.

Prepare is Ambience’s pre-visit workstream: the experience that helps clinicians get ready for an encounter before they walk into the room. Within Prepare, Chart Prep gives clinicians an organized, evidence-linked briefing of the patient’s longitudinal record, while Chart Chat lets them ask questions about the chart in natural language. This post focuses on Chart Prep and how Chorus, Ambience’s underlying context system, helps the AI investigate the longitudinal record, follow relevant clinical threads, and bring forward the information most likely to matter for the encounter.

A summary can be accurate and still leave the hard question unanswered

Consider a urologist preparing to see a patient with a recent multifocal stroke before a planned transurethral resection of the prostate (TURP). One of the key questions is the patient’s current antithrombotic plan.

The patient's records show that cardiology planned to change his medications if atrial fibrillation was detected. A later test confirmed atrial fibrillation, but subsequent records leave it unclear whether the medication change actually happened.

Each entry can be summarized faithfully. But the central question remains: was apixaban started, was clopidogrel stopped, and what is the right plan before surgery?

A system that simply compresses the notes can preserve every detail and still leave the clinician to reconcile the story. The preparation task is to connect the events, notice the contradiction, and make the unresolved question hard to miss. It should not guess at the answer, but instead show the clinician what the record supports, where the evidence conflicts, and what needs to be clarified.

That distinction is the difference between summarizing a chart and preparing someone to act on it.

Built for the minute before the visit

Clinicians do not have unlimited time to review a patient’s history. Chart Prep supports the short window before an encounter: orient the clinician, restore the relevant thread of care, highlight what has changed, and surface open questions while keeping the supporting evidence close at hand.

This ultimately gives clinicians a stronger starting point than a summary alone can provide. Rather than returning a smaller version of the chart, Chart Prep investigates the record in light of the upcoming visit. In the antithrombotic example, that means connecting the planned procedure to the medications relevant to bleeding risk, following the stroke-prevention plan through cardiology, checking the rhythm-monitor result, and comparing the intended medication change with the later reconciliation.

This is what it means to multiply a clinician’s capacity, not replace their judgment: take on more of the work of finding and organizing the relevant context, so the clinician can focus on interpreting it and deciding what to do next.

Why this takes more than putting the chart into a model

A patient’s longitudinal record is not one coherent document. It is a collection of notes, results, orders, medication records, and other events created at different times by different people and systems. The same fact may appear in several places. A plan may change. Two entries may conflict. And the most important clue may be in a different part of the record from the note that describes the upcoming visit.

Simply giving a model as much chart text as possible does not solve the problem. Long records can exceed a model’s context window, and even information that fits may be harder for the model to use reliably when the context is crowded. More importantly, the model still needs to know what to look for, which thread to follow, and when to inspect the underlying evidence.

Clinicians navigate records in stages. They orient themselves, follow a question through relevant encounters and results, open source material, and compare what different parts of the chart say. Ambience’s clinical team helps bring that practical understanding into product development: what a clinician would look for next, which contradictions deserve attention, and when the record is too ambiguous to support a firm conclusion. The system needs to support that kind of purposeful investigation, not just read a large bundle of text.

Chorus provides the underlying context system. It gives the agent tools to survey the record, follow a clinical thread, open underlying evidence, compare sources, and trace conclusions back to where they came from. Chorus maintains a shared, source-linked understanding of the patient that Ambience capabilities can use, rather than forcing each workflow to interpret the chart independently. The ambition is to turn more of the investigative work clinicians do into a system that can extend their capacity. This is not an AI pretending to be a doctor or substituting its judgment for theirs. It is AI built with clinical expertise that can do more of the preparatory work, so the physician can use their own judgment more effectively.

In the example above, the important finding does not live in one note. It emerges from the relationship between the cardiology plan, the positive rhythm-monitor result, the subsequent phone encounter, and the later medication reconciliation. The system has to find those pieces and connect them without treating the conflicting entries as if they agree.

To ensure confidence, Chart Prep shows clinicians the evidence behind its findings and flags when the record is incomplete or conflicting.

How do we know whether Chart Prep is preparing clinicians well?

Checking whether a generated statement is factually supported is necessary, but it is not enough. A supported fact can still distract from the visit, while an omitted fact can change what the clinician needs to do next.

So we evaluate preparation around several questions:

There is no single universal summary that is right for every patient and every visit. What matters depends on the patient’s history, the purpose of the encounter, and the clinician’s role. That is why evaluation starts with clinical judgment, rather than treating a generic summary as the answer key. We evaluate whether the AI identifies the information clinicians consider most important for the upcoming visit, and if it provides the right evidence behind each finding.

Board-certified clinicians review de-identified charts as they existed immediately before an upcoming visit and prepare as they would for care. They identify and rank the facts that matter for that encounter, bringing the same clinical priorities that shape the product into the evaluation itself. A second clinician independently reviews the case, and disagreements or omissions are adjudicated. This creates a clinician-established reference for what the system should find, what would be valuable to include, and which omissions could change the visit.

We then use automated reviewers to compare the assistant’s output against those references. The reviewers are calibrated against clinician judgments and used only after reaching an acceptable level of agreement. That lets us repeat evaluations across cases and product versions while keeping clinical judgment at the foundation. We also run each case multiple times because the quality of a single output does not tell us how consistently the system performs.

The evaluation is designed to measure more than whether the output sounds plausible. It tests whether the agent found the right information, connected it to the right evidence, and surfaced it in a way that supports preparation.

Improving the system, not just scoring it

We use these evaluations to identify where the system falls short and to guide improvements. When an important fact is missed, we examine the investigation trace to understand whether the agent failed to retrieve the evidence, failed to connect it to the relevant clinical thread, or found it but did not give it enough prominence. Clinical expertise helps us interpret why the miss matters in practice, while the trace helps us pinpoint what needs to change in the system. Those are different failure modes, and they call for different fixes.

On a benchmark of 46 de-identified patient cases, with three independent runs per case, the latest hill-climbing version of the context harness surfaced 92.2% of critical facts (95% confidence interval: 87.2% to 96.3%), compared with 78.7% for the production harness, with both versions scored against the same refrozen clinician references. Recall of all critical and high-importance facts increased from 66.3% to 72.3%.

We also examined the misses that remained. The share attributed to failure to retrieve the necessary evidence fell from 19.9% to 9.3%, a reported 53% reduction.

For this comparison, the generation model, patient cohort, clinician-established references, and evaluation suite were held constant. The principal change was the context harness built on Chorus. In other words, the improvement was not simply a change in the underlying model. It came from improving how the system investigated the chart.

These results show why the design of the context system matters: how an AI investigates a record can change how much of the clinically important information it finds. And because the evaluation helps us locate the source of a miss, it gives the team a concrete way to keep improving the product.

Preparation is the starting point for useful clinical AI

Before AI can meaningfully extend clinical work, it needs a sound understanding of the patient’s current state: what has changed, what matters for the decision ahead, and what remains unresolved.

That is the ambition behind Ambience Prepare and the broader context system Chorus supports: Not to give clinicians one more summary to read, but to use AI built with clinical expertise to take on more of the investigative work that goes into preparing for a visit, then make the evidence and uncertainty clear enough for the clinician to use.

In the case of the patient with a recent stroke who was preparing for prostate surgery, success would not mean the AI decides which medication the patient should take. Rather, it's about the clinician seeing how the plan may not have been reconciled, understanding why that question is still open, and addressing it before proceeding.

Our broader ambition is to multiply the impact of clinicians. We want to help them bring more of their expertise to each encounter by reducing the work of reconstructing the patient’s story. The AI won't replace the clinician's judgment, but it can make the judgment more efficient and effective by ensuring the right context is close at hand.