Where Vertical AI Can Win: Munjal Shah on Clinical Voice Agents and the Economics of Additional Care

1. Guest Background

Munjal Shah is the co-founder and CEO of Hippocratic AI, a company developing clinical voice agents that speak with patients over the telephone. Its customers include health systems, payers, and pharmaceutical companies. In this Exponential View interview, host Azeem Azhar asks him to connect the experience of building that business with a broader question: what can a specialized AI company offer as general-purpose models become more capable?

Shah approaches the question as a repeat entrepreneur with experience inside and outside healthcare. He describes Hippocratic as his fourth company, saying that his first two businesses were sold to Alibaba and Google. A personal health episode involving chest pains and an emergency-room visit after the second sale helped direct his next venture toward healthcare. That third company, he says, was less successful than he had hoped.

His interpretation of that experience shapes the episode. A compelling personal mission did not automatically produce a strong business. With Hippocratic, he wanted to combine a healthcare mission with a major change in AI capability, describing the venture as an AI company operating in healthcare. This background gives his argument practical relevance, while also establishing its perspective: the speaker is explaining the strategy of a company he leads. Its commercial and clinical claims should be read with that attribution intact.

2. What the Episode Covers

Shah presents Hippocratic as a business already operating at substantial scale. He reports approximately 300 employees, about $450 million raised, and a product launched around 21 months before the interview. He also cites 290 million patient interactions. Later, he distinguishes that figure from approximately 16 million to 17 million telephone calls. The distinction matters: interactions are not calls, and neither figure establishes the number of unique patients served. Volume provides operating context without independently demonstrating safety or clinical benefit.

The account of safety begins with authority inside the company. Shah says Minesh, its clinical and safety lead, can stop the company and bring safety concerns directly to the board without passing through the CEO. He also says Hippocratic withheld its product for its first 18 months until an internal benchmark indicated safety comparable to a human. The concrete governance mechanism is significant because it gives safety a route to interrupt commercial activity. The benchmark claim, however, is a company assessment; the conversation does not provide enough methodological detail to reproduce it.

The product is deliberately restricted. Shah describes a scope that excludes diagnosis and prescribing, along with categories including children under two, high-risk pregnancies, and mental health. Although he relates this scope to nursing work, these are the company's chosen boundaries and should not be treated as a general statement about every nurse's professional remit. Within those boundaries, one model conducts the conversation while 30 other models supervise. Shah claims this architecture benchmarks more safely than a single model, without presenting a fully reviewable comparison in the interview.

His next layer is output testing. According to Shah, roughly 7,000 US-licensed nurses participate through a testing platform, with license verification, simulated patient conversations, and error marking. Once in every 50 tests, a human nurse is placed on the other end to create a comparison. He reports spending $40 million and testing each new model and use case. This is a more specific validation account than simply listing training sources, although error severity, scenario selection, and the strength of the comparison remain insufficiently explained.

Monitoring continues after release. Shah describes listening to the first 1,000 calls for each new use case, calling patients back within an hour, and sampling 0.5% to 1% of all calls. He reports approximately 32,000 transfers to humans. Together, these measures describe an operating system for oversight and intervention. They do not, by themselves, reveal how many problems sampling misses or establish the overall clinical consequences of deployment.

3. Core Views: Reasoning, Examples, and Limits

Shah's broadest strategic claim is that the timing of a startup matters because opportunities are unevenly distributed across technology cycles. His forest-fire analogy contrasts finding a small opening beneath established companies with building after a platform shift has reopened the landscape. It explains his urgency around generative AI and his decision to reconnect healthcare with his AI background. The analogy is an interpretation of entrepreneurial experience, however, rather than evidence that businesses entering a new wave generally succeed. A favorable technological opening still leaves demand, execution, and competition unresolved.

His advice on founding teams follows a similar logic of improving the odds. Shah favors sharing ownership with people who can supply the capabilities needed to reach the first milestone, reducing dependence on later critical hires. The relevant tradeoff is between a larger individual stake and a team better equipped to make that stake valuable. For a company spanning clinical work and model engineering, the reasoning is readily understandable. It does not establish an optimal founder count or imply that adding partners without a clear capability need improves a business.

The more distinctive argument concerns the relationship between specialization and verification. Shah rejects the idea that knowing a model's training inputs establishes its safety. Actual behavior has to be tested. He then acknowledges a boundary to his own method: unrestricted general-purpose models admit too many possible inputs for the same approach to cover readily. A bounded agent with a specified prompt and job is more tractable. Verticalization therefore changes what a developer can meaningfully evaluate, as well as what it can sell. Greater testability still falls short of proving that real-world failure modes have been exhausted.

Open weights supply another part of his explanation. Shah says all 31 models are based on open models and have received fine-tuning, reinforcement learning, or other post-training. Access enables the company to alter the models themselves. He also estimates that running equivalent proprietary models for its architecture would cost $105 per hour, against a product price of about $9 to $10. These are workload-specific company estimates, not a universal comparison between open and proprietary systems. Their analytical value is to show how a proposed safety architecture can become inseparable from its economics: supervisors consume resources too.

Healthcare voice provides the clearest example of differentiation through simultaneous constraints. Shah argues that an agent needs roughly 500-millisecond latency while understanding clinically meaningful indirect language. A patient asked about shortness of breath might describe an elephant on their chest instead of answering yes or no. Fast turn-taking without that understanding would not satisfy the job. He reports changes to kernels, NVFP4 usage, quantization, and vLLM serving to prioritize response latency. The competitive proposition is an engineered combination of requirements that a general product may not prioritize together.

His restaurant-ordering comparison adds a purchasing argument. A hypothetical two-percentage-point advantage might matter differently when the outcome concerns clinical safety rather than an incorrect food order. This illustrates a possible willingness to pay; it is neither a measured safety improvement nor a demonstrated difference in mortality. A more concrete model failure reinforces the need for domain work: Shah says a company model offered telemedicine as an option for a Pap smear. He uses the example to argue that impressive reasoning can coexist with weak understanding of physical procedures. His announced additional $100 million in post-training is a spending intention, not evidence that the weakness has already been fixed. The episode also does not systematically establish his broader claim that successive models fail to improve such weaknesses.

Finally, Shah distinguishes three commercial approaches: Copilot assists existing workers, Autopilot performs existing work more cheaply, and Infinite Pilot supplies services that were previously uneconomic. He attributes almost all company success to the third. Replacement projects, he argues, confront existing performance standards, organizational resistance, integration, and retraining. Additional services can encounter less direct conflict. This is a useful explanation of where adoption may begin, grounded primarily in his company's experience. Lower organizational resistance does not establish that a service is valuable, nor that the same pattern will dominate every vertical market.

4. Learning and Application

An organization evaluating agents can start by separating assistance, replacement, and additional service proposals. Each needs a different business case. Replacement should be compared with the quality and output of the existing service, including integration and retraining. Additional service proposals should identify an actual unmet need and explain why addressing it matters. Shah's framework helps clarify those questions, but the absence of an existing workflow does not itself prove demand or eliminate the work required to support patients afterward.

His follow-up example offers a concrete starting point. A health system, he says, supplied 1,700 patients who had not completed recommended follow-up despite receiving texts, emails, and physical letters. Telephone conversations addressed questions that previous outreach had left unresolved, including why screening mattered when someone felt well and concerns about the equipment experience. A product team could use this example to investigate the specific obstacles between a reminder and an action. Interactive conversation may help where unanswered questions are the obstacle; the story does not establish that lack of information explains every missed examination.

Measurement should preserve the stages of the process. Shah reports 252 appointments scheduled, 180 examinations completed, nine significant findings, and three cancers. Those figures suggest separate measures for engagement, booking, completion, and follow-through. An implementation should also check whether clinical teams can accommodate additional examinations and manage abnormal results. Otherwise, successful outreach could simply move a bottleneck further into care delivery. This is an operational implication of the case, rather than an outcome the interview demonstrates.

Financial and health claims require separate scrutiny. Shah estimates $1,700 in campaign spending and $2 million to $3 million in revenue, describes three lives saved, and says one cancer was caught early enough for surgery without radiation or chemotherapy. The account lacks a controlled comparison, comprehensive cost accounting, and long-term outcome evidence. A buyer should therefore ask what the spending includes, how revenue is attributed, and what would probably have happened without the calls. Detecting cancer, generating revenue, and demonstrating an avoided death are different evidentiary achievements.

For technical deployment, the same discipline begins before model selection. Define permitted tasks, exclusions, and handoff conditions, then construct tests of actual outputs. Retest changes to models and use cases, and connect release decisions to monitoring after launch. Include supervising models, voice processing, and patient context in cost estimates. Use domain-specific failures to examine whether faster responses preserve the understanding the task requires. Narrower scope sacrifices some coverage, but makes product responsibilities easier to specify and assess.

Employment conclusions need comparable restraint. Shah says he is almost sure Hippocratic has not caused a nurse to lose a job, while acknowledging that he has not surveyed every customer. His account emphasizes additional services and redeployment; it cannot establish a sector-wide employment outcome. The practical way to use his proposed abundance revolution is to turn it into questions an institution can measure: which services expanded, how staff work changed, and what patients ultimately gained. That converts an optimistic vision into an evaluation program without treating the vision as an established result.

Source

• Original episode: Why vertical AI apps will win | Munjal Shah, Co-founder and CEO, Hippocratic AI