Shiv Rao on Building Abridge, Clinical Workflows and AI

Abridge co-founder and CEO Shiv Rao, a practicing cardiologist, speaks with Harry Stebbings about years of building before a market inflection, clinical workflows, AI models and company leadership. His descriptions of product capabilities, customer experiences and business performance are company claims, not independent clinical validation or medical advice.

Rao says he is motivated by the prospect of winning and creating a new product or market. Abridge began in 2018, and he describes several difficult years before broader interest in its approach. He believes a clear underlying thesis helped the team remain resilient while the timing of technology and customer adoption was uncertain.

The thesis, Rao says, was that conversations between clinicians and patients contain a valuable signal around which new healthcare workflows could be built. He was willing to change the product, features and go-to-market model, but not that central belief. He says the company would have had to start something entirely different if it abandoned the thesis.

Rao recalls raising a $5 million seed round at a $15 million pre-money valuation in 2018–19. He says finding a partner with whom he had chemistry mattered as well as founder-market fit. Those financing terms are his account of the round, not independently reviewed documents.

He describes following Union Square Ventures and investor Andy Weissman before meeting them, and interpreting their wide-ranging interest in music as a sign they might think creatively about healthcare. An introduction led to a pitch; Rao says he felt the relationship would work. The anecdote illustrates his approach to selecting a partner.

Rao says one of Abridge’s values is that people must experience good things to develop good taste. For him, taste combines pattern recognition with decisions that create an authentic product. He encourages the team to study research, user-interface patterns and work outside its own category rather than copying indiscriminately.

Asked what he has changed his mind about, Rao points to his assumptions about the company’s growth and possible scope. He says health-system partners are part of what he sees as a larger movement, while balancing ambition with gratitude. These are his reflections, not a measurement of customer outcomes.

Stebbings discloses that he invests in legal AI company Legora before asking about healthcare adoption. Rao says some physicians have told the company they would not join a health system without Abridge. This is a customer anecdote and his assessment of demand, not evidence that every health system or doctor has adopted the product.

Rao argues that US healthcare consists of multiple markets, not one uniform buyer. In his account, many practicing physicians are concentrated in large delivery systems and academic medical centers, so a company aiming to reach them must learn to serve large institutions. The workforce figures he cites are estimates in the interview.

He believes some healthcare startups remain in smaller customer segments too long rather than moving upmarket at the right time. Rao attributes Abridge’s opportunity around 2023 to several factors, including clinician burnout, financial pressure and improved AI. His cited burnout figures and causal explanation are not independently established in the conversation.

Rao believes being early in a vertical AI market can be valuable if the company survives long enough to adapt. He distinguishes companies founded before large language models from those built later, and says both the product and organization must keep evolving as the underlying technology changes.

Asked whether foundation-model companies could replace Abridge, Rao argues that vertical companies should find ways to work with model providers rather than treat them only as enemies. He sees differentiation in healthcare-specific workflows, integration and customer relationships. This is his competitive thesis, not proof that any company has an unassailable moat.

Rao says scale helps Abridge improve models and serve different clinician roles, care settings and spoken languages. He describes learning from user edits and argues that workflow integration is much more than a final layer on a model. He does not provide a controlled measure of the improvement or explain all data-use conditions in this exchange.

Rao reports that roughly 40% of model outputs within the product were generated by in-house models at the time, with the share changing as the company distills and fine-tunes models. The percentage and the prospect of a higher share are his description, not independently audited measurements.

For a bounded task, such as placing a particular item into a discrete field, Rao sees a case for an optimized in-house model. He suggests latency and cost can improve, while tasks requiring continuing quality gains may benefit from frontier models. The examples describe a design strategy, not evidence that any clinical task is error-free.

Rao says milliseconds matter inside clinical workflows and compares a good product to air conditioning that works quietly in the background. He describes helping a clinician approve an order or enter a visit diagnosis before a patient leaves. He emphasizes quality as well as speed but supplies no accuracy or safety results for those uses.

Rao says Abridge dynamically chooses between in-house and external models based on the user experience rather than ownership. He is not prioritizing cost over product quality now, though he expects costs may matter more later. His claim that the approach gives the company more control over its economics is a management assessment.

On when to optimize spending, Rao says it matters when cost limits the company’s ability to reach users or raise capital, not necessarily only when it lists publicly. He recounts investor Henry Kravis asking whether an IPO was needed; Rao says financing choices should serve the mission rather than become an end in themselves.

Stebbings asks whether enterprise healthcare data is difficult to prepare. Rao says obtaining, organizing and integrating data into workflows is complex and requires attention to security and compliance. He notes that behavioral-health information may need different treatment from other encounter data; he does not provide a full privacy or regulatory assessment.

Asked whether FDA review is required to sell an AI tool to enterprises, Rao says it was not required in Abridge’s case. That statement is specific to his company’s described situation; it does not establish the regulatory status of other products, features or intended uses.

Rao describes selecting a widely repeated workflow: the clinician-patient conversation. Abridge initially focused on generating a note, then explored adjacent work such as orders and billing. These are the founder’s descriptions of the product direction; he does not establish that these tasks can be safely or completely automated in every setting.

Rao says clinical documentation can affect how care is billed and that the company must take that responsibility seriously. He describes selling to different stakeholders, including clinical informatics, IT and finance leaders, each with a different priority. He views the note as a starting point for additional revenue-cycle workflows.

Stebbings asks about competition with electronic health record provider Epic. Rao says Epic has a note-related feature but describes Abridge as building an intelligence layer on top of records rather than trying to replace the record system. His differentiation and positive comparison of Epic with other systems are his opinions.

Reflecting on the pandemic, Rao says the company might have spent more time on research. He describes earlier work with models including BERT, BioBERT and T5, and says the team published research in 2021. He presents more research as a possible alternative, not a claim that a different strategy would certainly have succeeded.

Abridge initially built a consumer app intended to let patients ask permission to record a clinical conversation and receive a summary. Rao says the original vision included helping both doctor and patient. He considers the patient side central to the mission and says the company is returning to it.

Rao says the team struggled to find a consumer business model consistent with its trust goals and might have recognized the enterprise revenue opportunity sooner. His criticism of some consumer data business models is a general opinion, not evidence about any specific company. He says Abridge changed go-to-market and business model while keeping its underlying thesis.

On pricing, Rao says Abridge initially chose simple per-seat enterprise licensing. He describes competition with Microsoft’s Nuance and other providers but says his company now sees itself as a different type of offering. His characterizations of competitors and market positioning are self-interested assessments, not independent market-share data.

Rao recalls two doctors at Emory using “Abridge” as a verb and explaining that it meant more than producing a note. He says they associated it with reducing work and focusing on patients. The story is a small anecdote about user perception, not a quantified measure of clinical effectiveness or workload reduction.

He advises founders to focus primarily on their own product while understanding competitors’ positioning. He defines counter-positioning as building in a way an incumbent may find difficult to copy because it could affect an existing business. Whether that protection persists depends on market behavior and execution.

Rao says Abridge does not make money selling data because trust is central to healthcare. When Stebbings asks whether wider data sharing could benefit research, Rao responds that the question depends on partners and stakeholders. His statement is a company claim, not an independent audit of data practices.

For new features using data, Rao says the company seeks to “earn the right” through health-system conversations and contractual agreements. He says partners often know the roadmap and approve uses in advance. The interview does not detail every patient consent mechanism, data flow or contractual restriction.

Asked to choose between early access to frontier models and a top research team, Rao chooses people. He argues strong researchers and engineers can create their own models and tools, while temporary model access is not everything. He also describes the company’s mission as an attraction in a competitive hiring market.

Rao says the company aims to save clinicians time and reduce system costs, and describes a feature that suggests questions or diagnoses using clinical context and conversation. These are company goals and product claims. The interview gives no clinical accuracy, validation, safety or outcome data for the feature, which remains a support tool in his description.

Stebbings asks whether an AI that suggests questions and diagnoses should replace clinicians. Rao argues instead that clinicians have more work than they can fit into the day, citing a paper’s estimate, and describes assistants for chart preparation, notes and other administrative tasks. The paper and product effects are not evaluated in the interview.

Rao predicts some frequent tasks, including prescription refills, could increasingly be automated, while difficult care would still require clinician judgment. He characterizes refills as low-stakes, but risk depends on the medicine and patient; the discussion does not establish that an automated refill is safe or appropriate in any specific case.

Rao expects demand for healthcare to remain high and does not expect doctors to become unnecessary. Stebbings agrees that clinician demand will persist. This is their forecast, not proof of future staffing levels or a substitute for assessing the safety of clinical AI.

Asked about a reported forecast of a 24-fold increase in token use over five years, Rao declines to validate the exact number. He agrees with the direction, reasoning that lower inference costs and broader use can increase consumption. His view is a forecast, not measured future demand.

Rao says recruiting executives with strong judgment is especially hard when decisions and product cycles are moving quickly. He wants leaders who can draw on past patterns but also recognize when those patterns no longer fit. He sees this as a continuing challenge rather than a solved hiring process.

For Rao, “founder mode” means taking a focused tour of duty on an important problem, not micromanaging every person. He recalls investor Jensen Huang calling late at night in response to an email and discussing a challenge. Rao’s takeaway was to learn to embrace the tasks the job requires, including frequent travel.

Rao says his strengths also have limits and mentions managing HR as an area where he does not feel he operates at his best. Stebbings offers controversial views about HR and remote work; Rao describes Abridge’s current three-day office arrangement and says travel distances are part of its reasoning.

Rao values in-person collaboration but says Abridge does not impose a top-down seven-day or 996 schedule. He reports a team of around 450 people and thinks maintaining a high hiring bar requires continual attention. The headcount is his statement at the time, not an independently verified current figure.

Rao says he stopped personally meeting every new hire around or before the hundred-person stage. He sees no universal rule: trusted executives and hiring managers can carry the culture without the founder interviewing everyone. His approach depends on the people around him and the organization’s stage.

Rao looks for a deep sense of motivation and resilience, citing investor Josh Wolfe’s phrase about a chip on one’s shoulder. He distinguishes this from deliberately seeking damaged people. He believes some candidates have an unusually strong drive, although he acknowledges it can be difficult to assess.

Rao says pressure and dynamism matter to him and that coasting feels uncomfortable. Asked how to keep spending disciplined after a $300 million round, he advocates finance leaders who understand company goals and distinguish essential investments from discretionary expenses. The round size is his reported figure, not independently verified in the conversation.

He calls an earlier Series A1 his hardest financing, recalling several rejections while the consumer model had not yet reached an inflection. He credits existing and new investors for supporting the company, and describes the round’s pricing. These are his recollections, not a complete account of investor decisions.

Rao says leadership and parenting involve trade-offs and that his travel means missing some family events. He describes spending weekends with his wife and children and using perspective, purpose and gratitude to keep priorities in view. This is a personal arrangement, not a universal prescription for founders or parents.

He says Sunday time with his parents is non-negotiable and explains that aging parents are a major reason his family remains in Pittsburgh. He mentions his father’s heart failure as part of that personal context. The story illustrates how he balances travel with family commitments, not advice about treatment.

Stebbings questions whether San Francisco is too competitive for an early-stage company to hire. Rao disagrees, citing the concentration of ideas and talent. He says Abridge started in Pittsburgh and expanded its presence around the AI inflection in 2022–23, timing he considers right for the company.

In the quick-fire round, Rao says he has become more open to a flatter organization with fewer managers and more capable individual contributors. He believes AI tools could help people coordinate and take on more work. This is his vision of a possible operating model, not evidence that it will work in every company.

Rao credits early investors for backing him before he had a substantial track record. He thinks investor belief can strengthen founder confidence, but says the company must still execute. He names investor Elad Gil as someone he would call for a new company and credits him with lessons on focused corporate development.

Rao praises leaders he finds underappreciated and says he studies how they preserve culture and navigate markets. His examples are personal endorsements. He then returns to healthcare, describing providers, insurers and life-sciences companies as having incentives that are not always aligned with patients.

Rao hopes AI can support new business models that emphasize prevention while preserving access to experts for serious illness. He predicts early changes within three to five years. Those are aspirations and forecasts; the conversation does not establish that the proposed models will improve costs, access or clinical outcomes.