· Shiv Rao

Shiv Rao on Abridge: Building Healthcare AI in Founder Mode

Healthcare AI built with conversation-as-wedge, collaborating with frontier models, founder-mode principles across a hard down round and $300M scale.

founder-modehealthcare-aivertical-aifrontier-modelspricingenterprisetalentfundraising0% confidence

Why this is in the corpus

Shiv Rao articulates a vertical-AI playbook tested through a five-year wilderness and a brutal Series A down round: conversation-as-wedge, frontier-model collaboration, talent over arbitrage, trust as the speed-limit, and a flat IC-heavy org in the agent era. Concrete plays around segmentation, pricing, and model insourcing make this Tier 1 material.

Summary for skimmers

Abridge CEO Shiv Rao on building healthcare AI: holding the thesis (conversation as signal) while pivoting product/GTM/business model; segmenting a $5.3T market into integrated delivery networks for fast scale; collaborating not competing with frontier models; insourcing ~40% of model inference for latency; counter-positioning Epic and Microsoft Nuance; pricing per-seat for clarity; earning the right via partner consent; flat-org bets in the agent era.

Briefing

What survives the editorial filter

This page should feel like a smart colleague already listened for you and left only the operating logic worth keeping. Not everything said in the episode makes it through.

Trust signal

Direct episode extraction

Best used for

Decision-grade retrieval metadata not yet added for this episode.

Hold lightly

No explicit downgrade reason stored yet for this episode.

Principles

Durable claims that survive beyond the speaker's biography — each with explicit limits, transferability judgment, and evidence.

Principle

Chips on shoulders make chips in pockets

Hire for the chip on the shoulder.

Rao hires for people whose slope is fueled by something deep — the only sustainable engine for the resilience required by perpetual wartime.

Screen for the deep-stakes story behind a candidate's ambition.

Principle

Put the user first; cost discipline comes later

User-first beats stack-pride.

Rao only optimizes for cost when it gets in the way of impact. Henry Kravis directly disabused him of premature IPO thinking on the same principle.

Defer cost optimisation until it actively blocks impact; route every model choice through user outcome.

Principle

Win with the frontier, not against it

Treat frontier labs as wind, not waves.

Rao rejects the founder reflex to defend turf against foundation models — the win condition is structural collaboration plus proprietary depth they cannot reach.

Spend strategy cycles on how to ride frontier-model tailwinds before defending against them.

Principle

The market moves at the speed of trust

Trust is the rate-limiter on every revenue lever.

Rao refuses easy data monetisation because the industry's speed-limit is trust accumulation; Abridge compressed 15-20 years of trust-building into four.

Audit every potential revenue line against whether it compounds or burns trust.

Principle

Counter-position where the incumbent cannot follow

Counter-position so the incumbent's own P&L stops them.

Rao invokes Hamilton Helmer's Seven Powers directly: force the incumbent into a no-win.

Audit incumbents' revenue lines; build the product that, if copied, would destroy them.

Principle

Fall in love with whatever the job is

Adapt the founder, not the job.

Lesson from a midnight Jensen Huang call: redefine identity around what the company now needs.

When the company outgrows the founder's comfort zone, re-engineer the founder before re-engineering the role.

Principle

Pressure makes diamonds; coasting is the worst chapter

No pressure = ambition gap.

Rao treats absence of pressure as a defect to engineer out via bigger ambitions, not a reward.

When the company feels comfortable, 10x ambitions until pressure returns.

Principle

Hold the thesis, pivot everything else

Die on the hill of the thesis; pivot everything else.

Rao framed Abridge's five-year wilderness as a test of whether the thesis (healthcare is conversations) was strong enough to anchor relentless tactical pivots.

Separate what is sacred (thesis) from what is negotiable (product, GTM, pricing) and pivot loudly on the negotiable.

Principle

Conversation as wedge, not feature

Pick the signal, not the first deliverable, as your wedge.

Notes were the first product, but the wedge was the spoken doctor-patient conversation. That signal extends into orders, billing, decision support.

Ask: does owning this signal entitle me to expand into the next ten workflows without re-earning the relationship?

Principle

Reach down the stack to own destiny

Reach down the stack to own margin and performance destiny.

About 40% of Abridge model outputs are in-house; that figure flexes monthly as they distill new open-source models.

Identify the lowest layer of the stack you can defensibly own and migrate workloads to it.

Principle

You have to taste good things to have good taste

Taste is an exposure problem before it is a judgment problem.

Codified as an Abridge value pre-AI-taste-cycle; instructs engineers to read arxiv, study UX patterns, and live at the edge of culture.

Mandate cross-domain exposure as an operational input.

Principle

Founder Mode is tours of duty, not micromanagement

Founder Mode = selective tours of duty.

Rao reframes the Chesky concept as targeted intervention, not pervasive presence — coexisting with high-judgment executives.

Pick one tour of duty per quarter; trust execs everywhere else.

Frameworks

Reusable systems and operating models — including when they help and when they break.

Framework

Three-stakeholder enterprise sale: CMIO, CIO, CFO

Win the CMIO, CIO, and CFO together or none of them.

Healthcare's buying committee has three discrete decision-makers. Rao builds product features that align with money-flow specifically so the CFO is engaged from notes-onward (since notes become bills).

Build product roadmap with one feature line per stakeholder lens.

Framework

Binary vs continuous-improvement task split for model choice

Bell-ringable vs always-improving is the model-choice axis.

Rao's framework: for binary tasks, in-house specialised models optimize latency/cost and can be set-and-forget. For always-improving tasks (where users care about marginal improvement), ride frontier.

Tag every model call as binary or always-improving; let that drive build-vs-buy.

Signals

What appears to be shifting, for whom it matters, and what happens if you ignore it.

Signal

Verbing of brand by users as enterprise PMF

When users verb your brand unprompted, you have created a category.

Rao heard "abridge" used as a verb by multiple Emory clinicians on a panel and again at dinner — and the meaning was "did all these jobs for me and unburdened me."

Watch for user-driven verbing as the highest-fidelity PMF indicator.

Signal

LP meeting invitations as a power-shift indicator

LP meetings = power has flipped.

Rao quietly tracks LP-meeting count as a leading indicator of company status without making it a focus.

Treat LP meetings as a power barometer, not a chore.

Opportunities

Only included where there is a buyer, a real wedge, and a plausible revenue path — not vague idea theater.

Opportunity

AI-era business-model redesign of healthcare

AI's healthcare opportunity is business-model realignment, not science fiction in the clinic.

Rao expects realignment of misaligned providers, payers, and life-sciences companies via new models that incent prevention, materialising in the next 3-5 years.

Look for ventures restructuring economic alignment, not just deploying models.

Opportunity

Flat IC-heavy org as the agent-era operating model

Agents unlock flat super-IC orgs as a new operating model.

Rao explicitly says he changed his mind in the last 12 months and is moving toward a flatter Abridge with super-ICs leveraging agent tooling.

Treat flat super-IC org design as a strategic experiment, not a fad.

Lessons still worth keeping

Useful takeaways that did not fully clear the bar for durable principle status.

Lesson

The Series A-1 is a signal, not just a round

An A-1 is a public statement of weakness; price the social cost in.

Abridge raised at 2x on $100 from Winnington with no real numbers, pre-inflection — survival capital. Rao acknowledges the round itself communicated the moment in the lifecycle.

When a "-1" is the only path, plan the narrative for the next round before signing.

Lesson

Move to the AI epicentre when you are inflecting, not before

Relocate to the epicentre on the inflection, not the hope of one.

Abridge built in Pittsburgh on CMU DNA until 2022-23, then moved to SF as AI moment caught fire. Rao explicitly rejects relocating earlier.

Time geographic moves to inflection, not aspiration.

Lesson

Patient-side consumer pivot taught a real lesson but extracted a real cost

Some business models are unviable for trust-bound companies — recognise them earlier.

Abridge's early patient-facing app failed to find a clean business model. The lesson was less "wrong space" and more "recognise the cul-de-sac earlier and re-allocate to R&D."

Pre-vet pivots against trust-compatible business model availability before sinking cycles.

The Plays

Try these this week

Verb-first executable actions — each one tied to a stated outcome in the episode.

Become the latest AI vintage as fast as possible

Outcome: Re-platform to the newest vintage as soon as it ships.

Context: Rao identifies three vintages: post-transformer-pre-LLM, post-LLM-pre-agent, post-agent. Abridge rebuilt at each transition; the company and product evolved significantly each time.

depending on your vintage, you have to make sure that you become the latest variant as fast as you possibly can. And that might mean that your product evolves pretty significantly... we are a post transformer pre LLM and as soon as like the LLM moment like happened, we became that as fast as we possibly could
Shiv Rao
Cutover within 6-12 months of new vintage availability per
  1. 1

  2. 2

  3. 3

  4. 4

Before you start

  • · Modular product architecture
  • · Org willingness to refactor team structures
  • · Capital for parallel build

Earn the right before building features on partner data

Outcome: Paper data rights up front; never retrofit them.

Context: Abridge shares roadmap with partners in advance and papers data-leverage rights for future features into the contract, so trust scales with the product.

when we build new features that leverage data, we have a refrain in the company which is earn the right. So we always go back to our health system partners and oftentimes they know about our roadmap in advance and they've already blessed it and they've already, we've already papered this into the contract
Shiv Rao
At contract signing and on every quarterly roadmap update per
  1. 1

  2. 2

  3. 3

  4. 4

Before you start

  • · Enterprise sales motion already in place
  • · Legal/contracts function capable of bespoke clauses
  • · Sufficient product clarity to pre-commit roadmap themes

Run the in-house model swap when distillation beats frontier on latency or cost

Outcome: Insource the bell-ringable tasks; rent the frontier on open-ended ones.

Context: Rao runs a monthly review: which tasks are now solvable with a distilled fine-tuned open-source model? Swap them in. Frontier inference is preserved for tasks where "less imperfect every month" is the bar.

about 40% of our model outputs inside our product are generated by in-house models and that varies, You know, from month to month. Next month it might be 60% because we've distilled a new open source model and fine tuned it and gotten some feedback and we are convicted and we've just replaced a frontier model with this, this new in-house model
Shiv Rao
Monthly cadence of insourcing reviews per
  1. 1

  2. 2

  3. 3

  4. 4

  5. 5

Before you start

  • · User feedback/edit telemetry at scale
  • · ML team capable of distillation+fine-tune
  • · Rollback infra to frontier endpoints

Concentrate GTM on the integrated delivery networks first

Outcome: Skip down-market; time the YOLO shot up-market.

Context: Rao argues the trap in healthcare is staying down-market because the buyers (and clinicians) cluster in IDNs (Emory, Yale, UCSF) and payvider systems.

there's about a million doctors in the country, maybe 800,000 of them are actually practicing. The vast majority of them are concentrated in large care delivery systems... when they're concentrated there, you have to think about getting there as fast as you possibly can
Shiv Rao
12-24 months to land a flagship IDN, then expand per
  1. 1

  2. 2

  3. 3

  4. 4

Before you start

  • · A wedge that fits enterprise workflows
  • · Compliance posture (HiTrust, BAA-ready)
  • · Capacity to support multi-stakeholder buyer (CMIO/CIO/CFO)

Decision Moments

Actual decisions, real outcomes

Specific decisions narrated in the episode with their outcomes and transferable lessons.

Founded in 2018 around the thesis that the doctor-patient conversation is healthcare's most powerful signal — but technology (pre-LLM) and the enterprise market were not ready. Faced a 5-year wilderness through pandemic and tough Series A-1.

Did: Held the thesis as immutable while pivoting product order, GTM segmentation, and business model multiple times. Raised an A-1 at 2x on $100 from Winnington pre-inflection to survive. Built fine-tuned BERT/PEGASUS/T5 models pre-LLM so they could "become the latest variant" the instant the LLM moment landed in 2023.Outcome: When the LLM moment arrived in 2023, Abridge ran straight into a wide-open market. They raised a $300M round, became standard in IDNs, and reached the point where doctors verb the brand. Survived to be there when the sky opened.

Distinguishing the sacred (thesis) from the negotiable (product, GTM, business model) is what makes a 5-year wilderness walk survivable. Resilience is structural, not just emotional.

Part of an emerging decision pattern across multiple episodes

Pre-LLM, the Abridge team faced a fork: deepen R&D in the consumer-facing patient app or invest in being ready for a sudden enterprise market open.

Did: Built a direct-to-consumer patient app for recording doctor visits, then realized the only consumer business models in healthcare required selling sensitive private data — incompatible with the trust thesis. Pivoted hard back to enterprise + clinician side.Outcome: Lost cycles but learned that some business models are structurally incompatible with the trust posture required for enterprise healthcare. The realization came late but framed every subsequent pivot.

Pre-vet pivots against trust-compatible business model availability. Not every "valuable consumer product" can be monetised without burning the enterprise option.

Part of an emerging decision pattern across multiple episodes

Approached by Henry Kravis on whether Abridge should go public. Rao started "clumsily navigating an answer."

Did: Kravis interrupted with: you don't need to. The right answer is the mission — go public only if private markets cannot fund it. Rao internalized this and stopped treating going-public as a default ambition.Outcome: Removed IPO as a near-term forcing function on the company's strategy. Allowed continued private-market raises (300M last) and kept optionality on capital structure.

Cap-structure decisions should serve mission, not signal status. The right default for a mission-driven private company is to defer the public-markets question until it is the limiting constraint.

Part of an emerging decision pattern across multiple episodes

Tensions surfaced

Contradictions and trade-offs the episode raises — judgment calls a thoughtful operator has to navigate.

Tension

Per-seat clarity vs consumption-pricing upside

Simple-pricing-as-counter-positioning conflicts with maximal value capture.

Rao won early with per-seat against Microsoft's complex bundling — but watches other verticals where complex consumption pricing works. He has not resolved this tension.

Track whether your pricing simplicity remains a moat or starts to leave value uncaptured.

Tension

Automate the boring, but defend the hard

Automation appetite and human-judgment respect have to coexist.

Rao explicitly endorses automating prescription refills and primary-care basics; rejects "fire the doctors" thinking. Argues a healthcare-demand tsunami means more, not fewer, expert humans.

Build aggressive automation only where the stakes-and-frequency profile justifies it.

Corpus connection

Where this episode fits for retrieval

What kinds of decisions this briefing is best pulled into.

Primary decisions

  • strategic-bet
  • hire
  • pricing
  • market-entry