a16zHow Lassie Is Automating Healthcare Administration
CHAPTERS
- 0:00 – 1:17
AI is overhyped in Silicon Valley, underhyped on Main Street
The conversation opens with the core thesis: traditional software mostly digitized filing cabinets but didn’t actually do the work. The promise of AI is labor automation—especially impactful for underserved, non-technical small businesses.
- •Software historically stored information; humans still executed workflows
- •AI’s real wedge is doing tasks end-to-end, not just surfacing data
- •Main Street businesses face acute labor shortages and operational overload
- •Sets up why healthcare administration is a high-leverage target
- 1:17 – 3:46
The dentist paperwork shock: 200 hours/month of admin as the catalyst
Steijn recounts the pivotal moment: seeing a top-rated dentist spending massive time on claims and billing. What seemed like a solved problem remained painfully manual across many practices.
- •Dr. Kwan’s practice revealed extreme administrative burden
- •Manual claims submission and patient billing persist despite modern tools
- •Labor scarcity forces clinicians to do back-office work themselves
- •Early validation: this wasn’t an anomaly—other offices had the same pain
- 3:46 – 5:39
Pre-product immersion: earning trust and doing the billing by hand
Lassie’s founders embedded inside customer offices long before launching a product. The ease with which practices granted access signaled how desperate the need was—and how broken the status quo had become.
- •Founders positioned themselves as operators, not software vendors
- •Practices allowed deep access despite HIPAA/security concerns
- •Long, candid customer conversations confirmed urgent pain
- •Hands-on exposure created a granular understanding of real workflows
- 5:39 – 7:23
Building blocks first: context layer + tools before models were ready
Frédéric explains that early LLMs weren’t strong enough for complex reasoning, so Lassie focused on foundational infrastructure. As models improved, they could “swap in” better intelligence on top of existing context and tooling.
- •Automation requires context (historical records) and tools (actions)
- •Early models lacked reasoning; first use cases needed limited intelligence
- •Infrastructure-first approach created compounding leverage as models improved
- •Vision stayed constant: don’t build tools people must use—do the work
- 7:23 – 12:17
Why ‘software doing labor’ changes markets more than fintech bundling
Alex lays out an economic history of software and argues most digitization didn’t reduce headcount. AI expands markets by making software capable of executing workflows and charging for delivered work—not just licenses or storage.
- •From Sabre to Workday: software digitized records but didn’t automate jobs
- •AI can now edit/act on the data (e.g., collections, onboarding)
- •Fintech bundling (Toast) expanded software markets; labor automation expands more
- •Key insight: often you can’t even hire—AI fills supply-demand gaps
- 12:17 – 17:57
Adoption reality in SMB healthcare: relief from labor shortages, not job loss
The discussion reframes AI adoption as a response to staffing constraints and burnout. Dentists adopt quickly because the product directly removes late-night administrative work and improves quality of life.
- •Many practitioners quit/retire due to inability to hire admins
- •Lassie is positioned as ‘running the practice,’ not providing another dashboard
- •Economic model ties directly to labor budget; five-figure pricing can pencil
- •Customers value time back (family, coaching kids) as a primary outcome
- 17:57 – 22:17
Getting to 98% automation: humans-in-the-loop, then automating the humans away
Lassie achieved high automation by first doing the job themselves and iteratively removing manual steps. They target a threshold (95%+) before selling broadly, while learning from the long tail of edge cases.
- •Operational apprenticeship: doing the work was essential to product quality
- •SMBs lack staff to ‘use tools,’ so autonomy is required
- •Automation threshold approach: ship once it’s meaningfully hands-off
- •Continuous improvement via staff feedback on rare exceptions
- 22:17 – 30:03
Startup vs incumbent in the AI era: distribution still wins, but categories shift
Alex explains the TiVo problem and why owning distribution matters—incumbents can copy features. In AI, incumbents may innovate faster, but many “labor categories” had no true software incumbent in the first place.
- •TiVo lesson: innovation without customer control leads to weak outcomes
- •AI lowers the bar for incumbents to ship features faster
- •In many SMB workflows, the incumbent is human labor or outsourcing agencies
- •Defensibility comes from deep integrations, ontology, and workflow mastery
- 30:03 – 33:35
Onboarding as product: consumer-grade setup for non-technical, busy owners
A major differentiator is making deployment nearly self-serve—closer to Stripe/Rippling than enterprise consulting. The goal is minimal friction: connect bank accounts, systems of record, and insurance portals so the agent can act.
- •Main Street owners are time-poor; value must arrive fast or churn follows
- •Onboarding flow links bank, PMS/ERP, payer portals, and verifies business info
- •Under-the-hood configuration abstracts away operational complexity
- •Objective: ‘plug in and it works’ rather than training users on a tool
- 33:35 – 36:09
The master plan: from dentists to every doctor office—and eventually every SMB
Steijn describes a staged expansion strategy: dominate dental first, then adjacent healthcare verticals, then generalize across small businesses. The long-term vision is interoperating agents across businesses, consumers, and insurers.
- •Step 1: dental—large TAM, acute labor pain, repeatable workflows
- •Step 2: expand to other practice types with similar admin structures
- •Step 3: generalize to all SMBs with systems of record + customers + payments
- •End state: agents negotiating/transacting across the ecosystem
- 36:09 – 45:25
Hiring and building in 2026: steep-slope talent plus ‘AI-native’ execution
The founders share how prior lessons from Robinhood/Superhuman translate to today. They emphasize strict ICP control, fast time-to-value, and hiring people who can redesign functions around AI for a higher-output organization.
- •Maintain elite talent bar in engineering and sales; ambition still matters
- •Be strict on ICP to avoid painful mis-sells in labor-automation products
- •Measure onboarding like a playbook with timed milestones to core value
- •Screen for AI-native thinking across coding, finance, operations, and speed
- 45:25 – 50:34
What’s still hard technically: missing SOPs, slow learning, and workflow knowledge
Even large models lack domain-specific workflows and tacit office-manager knowledge. Lassie leans on historical operational data and product design to capture preferences and edge-case handling, while anticipating advances in smaller, faster-learning models.
- •Models don’t inherently know healthcare billing workflows or payer nuances
- •Tacit knowledge lives in office managers; not readily available online
- •Historical ERP/practice data helps infer workflows and exceptions
- •Excitement about smaller models that learn faster and can personalize
- 50:34 – 55:59
Digitization tailwinds and the ‘paper check’ problem: making automation possible
Steijn details how physical workflows (paper checks, mailed EOBs) block autonomy even with strong AI. Regulatory pushes toward digital payments plus Lassie’s conversion engine unlock end-to-end automation in a formerly paper-bound ecosystem.
- •Many practices still receive large volumes of payments via paper checks
- •Automation required converting payments/EOBs into digital, structured inputs
- •Regulatory mandates are accelerating electronic payments adoption
- •Product must handle permissioning and bank access safely for small offices
- 55:59 – 58:39
Reaching 160,000+ practices: building a new go-to-market for Main Street AI
The closing segment focuses on distribution: unlike enterprise AI, Lassie must reach hundreds of thousands of fragmented SMBs. They discuss mapping the market, identifying owners and systems, finding intent signals, and crafting messages that cut through noise.
- •SMB GTM requires scale outreach—not ‘a few steak dinners’ enterprise sales
- •Many owners aren’t in LinkedIn/Apollo-style databases, requiring new data approaches
- •Market mapping: location, ownership, tech stack, and intent signals
- •The hard part isn’t skepticism—it’s getting attention from overwhelmed operators
