Abridge: From Clinical Notes to Decisions During the Visit

The work that follows doctors home

The workday ends, but the notes do not. Clinicians catch up at home during pajama time, the name for after-hours documentation. Abridge describes doctors spending 10–20 hours a week on documentation—a measure of the burden, not measured time saved. Its internal Slack channel, Love Stories, collects clinician accounts of returning to family dinners or postponing early retirement.

Abridge started with clinical documentation. Its broader thesis is that patient–clinician conversations supply context for diagnoses, treatment, claims and payments. Chai describes his work on clinical decision support as bringing together patient context, payer guidelines and medical literature to help clinicians before, during and after visits.

The company describes three stages of value: save clinicians time, help health systems save or make money, and ultimately improve patient outcomes. It reports that its software is opened millions of times a week around patient encounters. That workflow presence offers a place to deliver decision support; improved patient outcomes remain the ambition.

Earn the interruption

Like Glean’s workplace information tools, Abridge depends on giving capable models relevant context. Healthcare changes both the scope and the stakes. Specialties and hospitals vary, but within a narrower domain than cross-industry enterprise search. An error such as recommending a medication to which a patient is allergic can be fatal, shaping both evaluation before deployment and gradual rollout to users.

Abridge’s ambient interaction model listens in the background so clinicians can focus on patients. The design aspiration resembles air conditioning: quietly improve the environment, then intervene when immediate action matters. An interruption must earn its place through useful, actionable information; noisy alerts risk being ignored.

A proposed pre-visit experience would summarize recent patient history and suggest topics relevant to the appointment. Preparing the doctor before entering the room could reduce manual chart review and interruptions during sensitive conversations. Other questions become valuable precisely while the patient is present.

Prior authorization, the insurer approval needed for a proposed service, makes that timing concrete. In a hypothetical example, a patient with knee pain receives an MRI order. Available context supports four of six requirements for a California Aetna plan; the system would prompt the doctor to ask about physical therapy and whether the pain has lasted more than six weeks. The envisioned outcome is approval before departure. These are illustrative requirements and a prospective outcome, not verified plan rules or demonstrated performance.

The mechanism is to resolve missing information while the patient is still present. It requires integration with the electronic health record (EHR), which holds patient context such as previous labs and imaging, and matching the patient to the correct payer policy. Policies vary by state and may live on websites or in unstructured 50-page PDFs. Assembling that context and meeting the accuracy requirement are potential competitive advantages. The ambition is to move weeks or months of subsequent work into the encounter itself.

Fit the encounter, the buyer and the patient

Abridge’s primary interfaces at the time of the conversation are mobile and desktop. Clinicians carry mobile devices between rooms and use desktop to finish notes or prepare for the next day. Nursing exposes a different constraint: starting a session can take longer than a brief check-in. Visits lasting 30 seconds or two minutes motivate discussions with in-room device companies about capturing more context through always-on devices. Glasses that overlay imaging are a separate, speculative direction outside the near-term roadmap.

Product design distinguishes three constituencies: - Buyers: Health-system medical information, finance and information-technology leaders need security, compliance and tangible financial value. - Users: Clinicians need an experience that fits their work and reduces its burden. - Patients: People downstream of the purchase experience the consequences for their care.

For finance leaders, time savings alone are insufficient. Abridge points to more compliant documentation and fewer billing-team queries as mechanisms for financial value, without quantifying a dollar return. Insurer collaboration also has a specific boundary: payers would receive information needed for authorization, not raw Abridge data.

Balance quality, latency and cost during the conversation

Authorization reasoning spans combinations of procedures, payer policies and sometimes health systems. It must be accurate enough to avoid unhelpful alerts, fast enough to guide the ongoing conversation, and affordable. An intermediate representation for payer policies—a form the system could reason over more readily—is floated as one approach. Its structure and deployment status are unspecified.

Abridge uses a constellation of models, assigning different models to different parts of the product. Proprietary development must offer higher quality, or comparable quality at lower cost and latency. The company says patient–provider conversations supply clinical reasoning traces that improve transcription, separation of speakers in transcripts, and note generation, with efficiency advantages it does not quantify. Improvements in third-party models’ healthcare knowledge and ability to carry out tasks can benefit other parts of the system.

One architectural opportunity is to treat the EHR as a filesystem. A patient’s record can exceed the model’s context window. An agent could selectively read and manipulate external information rather than require the whole record in one prompt. This is a proposed direction for making that context usable.

The existing experience is batch-based, with no interval specified. Prototypes explore triggering models or agent workflows at relevant moments to shorten the feedback loop. Output remains primarily text: a speaking agent could become an unwanted third voice in the room. Fuller voice interaction depends on better quality and greater clinician and patient comfort.

Personalization changes what good output means

A clinician reportedly refused to adopt the tool unless it put two spaces between sentences. The detail captures how personal a clinical note can be: it represents the doctor’s work. Abridge organizes personalization at three levels: - Individual clinician: Bullets or paragraphs, concise or comprehensive coverage, preferred phrases and templates—all while preserving clinical accuracy. - Medical specialty: Dermatology and primary-care notes have different completeness, compliance and billability criteria, changing both onboarding and evaluation. - Health system: Hospitals want their established guidelines embedded in decision support before, during and after visits.

Institutional differences can look small while changing a decision. Two systems may both have specialist-referral guidelines but require different conditions before referral. Clinician edits offer another source of context, feeding improvements in personalization over time.

The direction under exploration is persistent memory separate from model weights. Rapid model replacement could make preferences encoded in weights disposable work. A separate store would let a background memory sub-agent identify clinician actions worth retaining, while background jobs consolidate information from note edits and conversations. This remains an architectural direction rather than a fully specified production system.

Evaluation requires clinical judgment and operational capacity

Abridge defines evaluation criteria when starting a feature. Clinical safety is the baseline; note quality also includes style, completeness and support for billing. In-house clinicians first inspect outputs directly through LFD, the company’s emphatic shorthand for looking at the data. Large language model judges then assess separate quality dimensions, calibrated against annotated data and human evaluators. Major changes receive additional internal or third-party review according to their stakes.

The evaluator must match the output’s purpose: - Specialty-specific notes may require specialist clinicians to judge their quality. - Billing-oriented documentation may require healthcare coding teams to assess whether the note represents the care delivered and sufficiently justifies the prospective charge.

Building that evaluator network and planning capacity across specialties require substantial operational investment. Moving evaluation from months toward weeks or days is therefore a staffing and coordination problem as well as a machine-learning problem.

A large offline evaluation set may still have the wrong distribution. Progressive rollout tests usefulness in real workflows and feeds new cases into evaluation. Abridge reports moving customers from historically quarterly or twice-yearly vendor releases to monthly releases. Some customers also participate in gradual development releases before general availability. Their trust enables faster learning while the bar for experimentation remains high.

Learning from clinical data introduces another model to evaluate. Abridge builds models that remove indicators of protected health information (PHI) from transcripts before permitted training or evaluation use. The upstream de-identification model must itself be validated: otherwise, the pipeline stacks probabilistic systems without confidence in the first step. Abridge describes anonymization as one-way, with partner contracts governing access, retention and data use.

Scale makes efficiency and interoperability decisive

Abridge later describes its dataset as 100 million conversations, though an earlier estimate varies. At that reported scale, model cost and token consumption become central concerns. Its post-training team targets compute and token efficiency, especially where little quality headroom remains and off-the-shelf models may not optimize in the desired direction. The development sequence is practical: establish a useful product, then use accumulated data to improve its efficiency.

The platform ambition follows from the different questions one conversation could answer: - Doctors: Does the documentation fully represent the care delivered? - Patients: What happened, and what are the next steps? - Payers: Was the care appropriate? - Pharmaceutical companies: What might explain drug use, or identify a clinical-trial candidate? Abridge envisions consolidating separate systems serving these needs. Greater efficiency and better outcomes are intended benefits of that platform.

Another proposed workflow starts when a lab value changes. A background agent could combine that event with patient context, suggest a next step and flag it to a clinician who remains in the loop. The objective is lower latency to care, alongside better care quality and lower cost. Direct patient-facing connections are further away.

These possibilities depend on deep EHR integration. Abridge credits close partnerships with enabling data exchange through APIs that were not ready out of the box. Reading and writing information in the correct places is a prerequisite for adoption. The company recounts clinicians rejecting products over two additional clicks: with back-to-back 15-minute appointments and unfinished documentation, workflow friction can erase a feature’s value.

Reliability changes the team and the bottlenecks

Healthcare’s constraints also motivate optimism. One participant sees government support for interoperability and the Food and Drug Administration’s January clinical-decision-support guidance as favorable developments relative to its 2022 guidance. Another argues that healthcare may force difficult reliability problems to be solved earlier because products cannot ship without addressing them. Zero-error evaluation is an ambition here, not a measured result.

Chai’s latency advice starts with identifying the actual bottleneck. When the model dominates, post-training can target efficiency. Routing offers another approach: a cheap, fast model triages work before handing it to a larger model for greater capability. These techniques help make products useful now, even if changes in model capabilities make them insufficient or less relevant over a five-year horizon.

Clinician scientists bring clinical judgment inside product teams. Typically trained as physicians, their technical abilities range from full-stack engineering to resourceful prompting. They assess clinical usefulness and help define evaluation criteria that nonclinical product and engineering staff lack the expertise to set alone.

Abridge is also making evaluation sizing more systematic: deciding when several hundred offline responses suffice and when several thousand are needed to catch long-tail failures before production. Internal and third-party evaluators supply offline assessments; consenting customers contribute ratings and preferences between outputs.

What survives better models

Chai credits Glean’s success to a concrete information-finding problem and strong search foundations, including a team experienced in building search at Google. His lesson is that poor execution can make a useful category look unpromising. Putting Glean’s context layer to work as AI changes remains both an opportunity and an adaptation challenge.

The infrastructure discussion separates potentially temporary work from more durable needs: - Model-specific scaffolding may expire. Earlier agent development required a custom domain-specific language to simplify tasks for weaker models. Better computer use and sandboxed coding shift attention toward the context and tools agents need. - Event-driven systems remain relevant. Kafka, Temporal and sockets are cited as familiar technologies for coordinating real-time conversational workflows. - Collaboration still needs conflict handling. Google Docs-style collaboration offers precedents for managing conflicting updates in multi-agent systems, with speed and scalability still essential.

Cheap prototypes make written judgment more valuable

Abridge’s product-development position challenges the idea that prototypes make product requirements documents obsolete. Written clarity supplies context for both teams and AI systems, even when the format becomes Markdown. A prototype struggles to answer several consequential questions: - Data use: What can the product do with the available data? - Investment: Is this the right problem to spend time solving? - Differentiation: Why should the company win if competitors build the same feature? - Implementation: What must each customer do to use it?

The strategic test is deliberately demanding: if 20 competitors build the feature too, what remains distinctive? A host argues that building first can itself be the way to learn. The response ties the higher investment bar to the expense of getting software into health systems. Prototypes remain valuable for showing clinicians alternatives and increasing fidelity, while implementation, security, compliance and edge cases still need written treatment.

The prior-authorization ambition returns as a measure of scope: collapse a process described as 45 days and roughly 20 touchpoints into one touchpoint. That is the target, not an achieved reduction. Despite the gap between a weekend demo and software working in large health systems, a participant describes this as the fastest development of their career when accounting for complexity and scale. Upfront thought concentrates speed on the right problems. Small changes may still justify shipping a prototype directly; judgment should adjust as costs and benefits change.

The tooling discussion offers a glimpse of that pace: Abridge reports heavy use of Claude Code, with Cursor also available. One anecdote describes an engineer running about six Claude instances and using mobile remote control.

The closing invitation points to Abridge’s technical white papers, including work on reducing hallucinations, and its AbridgeHQ account for further technical material. The emphasis returns to evaluation: rigor begins by defining what constitutes a problem. An AI-and-healthcare open house is also announced, without a date or event link.