Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board
Bret Taylor has navigated nearly every major Silicon Valley platform shift—Google Maps, Facebook's Like button, co-CEO of Salesforce, chairman of OpenAI. Now, as founder and CEO of Sierra, he's building AI agents for customer experience and watching the largest companies in the world absorb this technology in real time. But as Sierra crosses $165 million in ARR in under two years, Taylor sees something surprising: the real barrier to AI productivity gains isn't model capability. It's that companies are still organized around departments instead of processes, still building tools for humans instead of outcomes, still trying to preserve org charts designed for a pre-agent world. The question hanging over every board room is no longer whether AI will reshape software—it's whether incumbents can move fast enough before new architectures replace them entirely.
Pontos-chave
Outcome-based pricing—charging only when an AI agent fully resolves a case or completes a sale—is the CPC ads moment for enterprise software, fundamentally realigning incentives between vendors and customers.
The atomic unit of AI productivity is a process, not a person; companies still organized by department (legal, finance, IT) will struggle to capture value until they reorganize around end-to-end workflows like supplier onboarding or contract negotiation.
Coding agents have progressed so rapidly in four months that most Silicon Valley companies will stop writing code by hand by year-end 2026—a state change that seemed impossible in late 2025.
High-agency generalists who deeply understand customers but weren't the strongest individual engineers are suddenly the most valuable people in tech companies, empowered by AI exoskeletons to ship without bottlenecks.
The software industry's $300+ billion valuation correction reflects rational uncertainty, not irrationality: if recurring revenue isn't truly recurring in an agentic world, the entire SaaS discounted cash flow model breaks down.
Em resumo
AI agents will become the primary digital interface for most businesses within two years, but the companies that win won't be those with the best models—they'll be the ones brave enough to reorganize around end-to-end processes, charge for outcomes instead of seats, and empower high-agency generalists with exoskeletons of code.
The Janky Genius of OpenClaw
A chaotic open-source project reveals AI memory beats polish.
OpenClaw—the semi-rogue, three-name-changes-in-three-days project running on Mac minis and chatting over WhatsApp—shouldn't work as well as it does. Yet Taylor calls it «probably the first broad hobbyist use of AI» that actually delivers. The reason? Memory. While polished consumer AI apps like ChatGPT and Gemini still open as blank slates with no memory, OpenClaw writes notes to a markdown file like the movie Memento, creating continuity across sessions.
The technical insight is profound: code repositories were accidentally optimized for AI. They're textual, structured, version-controlled, and filled with feedback loops—compile errors, unit tests, code reviews. «It's almost designed for a robot,» Taylor observes. That's why coding agents have transformed in just four months. The same file-system approach—treating life as source code—may be the most efficient harness for general-purpose agents, at least in the near term.
Taylor believes the messiness is a feature. Random access memory (vector databases) requires knowing what to look for; real memory is a mix of context and retrieval. The markdown-file approach, however crude, produces more useful agents than fancier architectures. «The idea that there's a directory of just everything you've ever done is actually maybe more useful to an AI than people think.»
From Code Craftsmanship to Outcome Obsession
Taylor is forcing himself to stop caring about code elegance.
“I have a hard time not caring. I don't care about the assembly language produced by the compiler. Why should I care about the code? I care about correctness, I care about robustness, and I think I know intellectually, I don't need to look at how the compiler unrolled this loop to verify its elegance and correctness. Yet somehow I feel that way about code.”
What Sierra Does and How Fast It's Growing
AI customer service agents hit $165M ARR in nine quarters.
The Hidden Architecture of AI Adoption
Companies fail because they optimize departments, not end-to-end processes.
The Hidden Architecture of AI Adoption
The barrier to AI productivity isn't technology—it's organizational structure. Taylor argues that AI operates on processes (onboarding a supplier, negotiating a contract), not people. But companies are organized by department: legal does contracts, procurement negotiates, IT onboards systems. No one owns the 17-day supplier onboarding process end-to-end. «We ship our org charts,» Taylor says. Until companies reorganize around workflows with clear KPIs and accountable owners, they won't capture AI's value. The flower shop test makes this concrete: give a local florist all the AI in the world, and someone still clips stems, arranges bouquets, and thanks customers. Digital workflows—finance, legal, compliance—can absorb intelligence efficiently. Physical ones can't. Not yet.
Why Outcome-Based Pricing Changes Everything
Charging for results, not usage, realigns software vendor incentives completely.
The Agent as UI: Rethinking Software's Front Door
Web forms and dashboards were a moment in time; agents are the future interface.
What if navigating websites and filling out forms was just «a bit of a moment in time,» like faxing emails over telephone lines? Taylor believes a company's AI agent—singular, branded, conversational—will become «the vast majority of their digital interactions.» Not because websites disappear, but because agents work across every channel: phone, WhatsApp, website chat, mobile app. They're future-proof because they're fundamentally conversational.
The shift goes deeper than modality. Traditional software built elegant UIs for humans (Stripe's dashboard is famously beautiful). But the ideal agent harness won't be elegant—it'll be optimized for context density and multi-step procedures. Taylor draws the parallel to coding agents: the harness you write for a software agent is radically different from UI design. «I wonder if the web application of the future will actually be... an agent harness,» he muses. Not just APIs, but the instruction manual for the Unix commands that power your business.
Stripe already built the APIs for agentic commerce a decade ago, for the failed «social shopping» wave (buying on Instagram, buying on Twitter). Patrick Collison has wanted to SSH into Stripe accounts for years—«tail the payments log, grep, pipe.» Now they're building it, because terminal-centric interfaces give agents leverage without requiring vector databases. The question isn't whether this happens. It's whether companies build agent-native harnesses before desktop computer-use agents just learn to click around legacy software. Dario Amodei's race: will services make themselves accessible to agents, or will agents just get better at using Chrome?
Why the SaaS Correction Was Rational (But Overblown)
The Rise of the Hyper-Generalist
High-agency people who care deeply are suddenly the most valuable employees.
The Old Constraint Generalists with great taste and customer understanding but weaker technical depth got sidelined as companies grew and specialized. There wasn't a clear role for them—they weren't the deepest engineer or the best designer.
The Exoskeleton Arrives AI agents give high-agency generalists an exoskeleton. They always had ideas about serving customers better; now they can ship without bottlenecks. Work ethic becomes addictive because «you can do so much»—people work harder because the marginal return on effort explodes.
The New Org Chart Taylor predicts flatter organizations built around these empowered individuals. The canonical Silicon Valley template—engineering/product/design ratios, go-to-market pipelines—was optimized for human labor. In an agentic world, care and taste matter more than specialization.
The Unsolved Problem What role do these people fit into? «Product engineer» means something else. «Minister without a portfolio» isn't an org chart box. Silicon Valley needs to invent a new job category for the product-designer-PM-engineer hybrid who ships with agency.
Lessons from the OpenAI Board
Fiduciary duty to a mission, not shareholders, changes everything.
Taylor joined the OpenAI board after the Sam Altman firing crisis, brought in as the mediator both sides agreed upon. It's his first not-for-profit board, and the difference is profound: «Your sole duty is to ensure that artificial general intelligence benefits humanity.» Not maximize shareholder value. Not hit revenue targets. Benefit humanity. That clarity reshapes every decision.
The board-building process was unprecedented—going from three members to a full slate almost from scratch. Taylor and the remaining directors had to think holistically: How do you represent safety? Economic impact? Infrastructure expertise? The not-for-profit mission? «Normally you add one board member at a time. This one was like, 'Do you have a bulk rate?'»
The research itself is «very inspiring.» It's easy to grow cynical watching model leaderboard races, but walking into board meetings where every researcher is trying to make safe AGI makes cynicism impossible. Taylor had never been affiliated with a true research lab before. The experience has been clarifying: fiduciary duty to a mission, not a balance sheet, is «really different» from every other board he's served on.
Predictions for 2026
Títulos mencionados
Pessoas
Glossário
Aviso: Este é um resumo gerado por IA de um vídeo do YouTube para fins educacionais e de referência. Não constitui aconselhamento de investimento, financeiro ou jurídico. Verifique sempre as informações com as fontes originais antes de tomar decisões. O TubeReads não é afiliado ao criador do conteúdo.