a16z’s Julie Yoo Says the Next AI Companies Will Sell Accountability, Not Intelligence

Frontier model progress is putting pressure on one of AI startups’ most common selling points: being smarter than everyone else. As OpenAI, Anthropic and Google keep pushing model capabilities forward, a16z General Partner Julie Yoo argues that startups need a different moat. Her answer is a framework she calls both AI-native and AI-proof. AI-native, in Yoo’s definition, means a company can use the best available models and turn that efficiency into real financial performance. AI-proof means something harder: even if GPT- or Claude-level models keep improving, the company is not easy for the model vendor to replace. In healthcare, Yoo says, the scarce product is shifting away from intelligence and automation alone. The real bottleneck is accountability. That point matters because the fastest-growing uses of AI in healthcare already fall into two buckets. One is intelligence: diagnosis support, clinical decision support and information gathering. The other is automation: charting, billing, scheduling and back-office work. Both are also the areas where frontier models are improving the fastest. If the core value is simply “our AI is smarter” or “our AI saves labor,” that defense weakens over time. Yoo says the stack changes from intelligence plus automation to intelligence plus automation plus accountability. Her first example is AI-native clinical services. Instead of selling software to hospitals, a company becomes the care provider itself. AI can cut labor and administrative costs and let a team serve more patients, expand across states, or move into new specialties more quickly. But the hard part is not the model. It is licensure, payer credentialing, malpractice insurance, referral agreements and actual offline care delivery. Even lower-regulation adjacent services still need safety testing, physician oversight and workflow integration. That is why GPT or Claude becoming ten times better would not suddenly make OpenAI or Anthropic a multi-state medical provider. Yoo’s second model is a risk-bearing entity. Here, a company does not just reduce costs for insurers or providers; it takes on the financial downside if cost control fails. A business could underwrite a patient population’s healthcare spend, using AI to automate much of the administrative machinery behind risk management. But a full-stack risk taker still needs reserves, state licenses, provider networks and ongoing management of medical spend. She says this idea is not limited to healthcare. A pure software company could also take on financial risk by charging only when a transaction closes or when a customer actually gets paid. In that setup, the company is selling outcomes, not software. That leads to a broader business-model shift. Traditional SaaS charges by seat, account or usage. But if AI agents can replace large amounts of human work, seat-based pricing becomes less durable. Outcome-based pricing may be the more AI-proof model: a company does not sell a tool to improve medical collections; it collects the money for the hospital and takes a cut of what it recovers. Yoo’s third example is FDA-regulated products, including diagnostics and drugs. AI can automate a lot of the cognition and operations around drug discovery, experimental design, data analysis and clinical workflows. But the real world still does not disappear. Companies still have to manufacture products, run years of human clinical trials and win FDA approval before they can claim a diagnosis or therapy works. She also points to data generation as another moat. Biology is still understood only at a very early stage, which means AI companies need more than public data. They need ongoing experimentation, new biological data and large validation loops. That makes proprietary data generation a barrier model companies cannot simply copy. Yoo’s bottom line is blunt: as models get smarter, intelligence gets cheaper. What becomes scarce is the responsibility that sits outside the model itself.