Mar 10, 2026 · 1h 41m · cheeky-pint
Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth conversation, Stripe co-founder John Collison sits down with Sierra founder and OpenAI board chairman Bret Taylor to explore the architectural principles, economic transformations, and organizational shifts driven by autonomous AI agents. Taylor details Sierra's journey framework and outcome-based pricing, critiques brittle multi-agent architectures, and shares firsthand governance insights from leading boards at Twitter and OpenAI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. John holds 20.8% of the talking time here. How this is scored →
John as informed peer 5.2 Guest teaching 5.2 Guest disagreement 1.9 John pushing back 1.7
speaking balance: gold is John, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:20:54 Taylor rejects department-level AI productivity framing
Taylor firmly interrupts Collison's inquiry about lagging white-collar departmental productivity, insisting that treating departments rather than cross-functional processes as the atomic unit of automation is fundamentally flawed.
Hardest push from John ▶ 1:09:25 Collison asks if Sierra is inherently short AGICollison directly challenges Taylor's long-term business moat by asking if Sierra is fundamentally short AGI given that frontier model labs continually swallow application-level scaffolding.
Biggest teaching moment ▶ 17:15 Taylor on legacy PSTN telephony trumping new protocolsTaylor educates Collison by revealing how real-world healthcare payers and providers already deploy AI agents that converse with each other in natural English over century-old telephone rails rather than waiting for modern API standards.
John holds their own ▶ 58:00 Collison explains the dissolution of database moatsCollison demonstrates deep technical command of SaaS architectures, articulating how LLMs remove the historic need for centralized, homogeneous databases by effortlessly normalizing unstructured inputs across distributed endpoints.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | John as informed peer | Guest teaching | Guest disagreement | John pushing back | Why |
|---|---|---|---|---|---|---|
| Why Coding Agents Excel and Harness Engineering Evolution | 5 | 6 | 2 | 1 | Taylor explains why coding agents have advanced rapidly due to structured text repos and immediate compiler feedback loops, contrasting them with unstructured general knowledge work. Collison contributes relevant observations on Unix tools and terminal ergonomics, keeping the exchange technical and collaborative. | |
| Artisanal Code Versus Emotional Detachment in AI Engineering | 4 | 3 | 1 | 1 | Collison inquires about Taylor's personal engineering habits in the age of AI coding. Taylor reflects candidly on the emotional challenge of detaching from hand-crafted, artisanal code while recognizing that modern software craftsmanship is evolving toward higher-level specifications. | |
| Model Context Protocol Critique and Context Rich Architectures | 4 | 6 | 4 | 1 | Taylor pushes back on the industry hype surrounding Anthropic's Model Context Protocol (MCP), arguing that over-architected sub-agent hierarchies strip necessary conversational context. He favors broader, file-based context models like OpenClaw over fragmented API servers. | |
| Command Line Ergonomics and the Evolution of SaaS Harnesses | 6 | 4 | 1 | 2 | Collison draws on internal Stripe product experiments, explaining why SSH-style terminal access into SaaS accounts is resurging. Taylor builds on this by proposing that future enterprise SaaS platforms will expose comprehensive 'agent harnesses' rather than traditional human dashboards or sparse REST APIs. | |
| Computer Use Automation Versus Plain English Telephony Rails | 5 | 6 | 3 | 2 | Collison cites Dario Amodei's thesis on computer use overtaking bespoke APIs. Taylor counters with a real-world counterintuitive example from healthcare, where AI agents communicate over legacy PSTN telephone lines in plain English rather than needing complex modern protocol integrations. | |
| Sierra’s Hypergrowth and Transforming Support Into Growth Centers | 5 | 6 | 1 | 1 | Taylor details Sierra's rapid revenue trajectory ($165M ARR) and articulates how reducing the marginal cost of customer interactions from $20 to pennies turns traditional cost centers into high-ROI retention and expansion channels. | |
| The Evolution of Digital Interfaces and Screen Independence | 5 | 4 | 2 | 2 | Collison queries whether web forms and browsing are transient historical artifacts like fax machines. Taylor provides historical perspective across computing eras, predicting that conversational voice/chat agents will become primary digital front doors while visual screens recede into specific niches. | |
| Automation Economics, Escalation Handling, and Market Shuffling Windows | 5 | 6 | 2 | 1 | Taylor explains the counterintuitive dynamics of deploying AI support: human agent handle times increase because routine queries are automated (70-90%), leaving complex cases. He stresses that AI capabilities will quickly become table-stakes industry imperatives rather than permanent competitive moats. | |
| Journey Frameworks, Foundational Knowledge, and Supervisor Architecture | 5 | 7 | 1 | 1 | Taylor breaks down Sierra's architectural approach to enterprise reliability: defining declarative journeys and utilizing multi-model supervisor architectures to audit reasoning chains and eliminate hallucinations for prominent enterprise brands. | |
| Building Applied AI Startups Around Ephemeral Infrastructure | 6 | 5 | 1 | 2 | Collison highlights the organizational friction of engineering teams maintaining custom code that underlying frontier models will inevitably commoditize. Taylor agrees, emphasizing that applied AI startups must embrace ephemeral scaffolding and compete on workflow product velocity. | |
| Software Market Valuation Uncertainty and Systems of Process | 7 | 6 | 2 | 3 | Collison and Taylor analyze public market SaaS multiple contractions. Taylor distinguishes between resilient general ledgers and vulnerable workflow systems of record, arguing that value is shifting from static databases toward active, encoded process execution. | |
| Stripe Sessions Conference Announcement and Registration | 6 | 5 | 3 | 2 | Following an ad read for Stripe Sessions, Taylor clarifies the distinction between usage-based token pricing and outcome-based pricing, arguing that aligning software revenue with business resolution rates creates superior incentives and accountability. | |
| Applied AI Longevity in an Approaching AGI World | 6 | 5 | 2 | 3 | Collison asks whether Sierra is inherently 'short AGI' if foundation models absorb application layers. Taylor mounts a defense of applied AI, noting that enterprise departmental alignment, procurement nuances, and bespoke workflow integration will remain defensible moats even under advanced models. | |
| Process-Driven AI Productivity and White-Collar Workflow Transformation | 6 | 7 | 4 | 3 | Collison presses Taylor on why white-collar productivity gains remain elusive outside of software engineering. Taylor reframes the problem, arguing that companies erroneously apply AI to generic departmental silos instead of decomposing cross-functional end-to-end business processes. | |
| Post-AI Startup Structure and the Ascendancy of High-Agency Generalists | 6 | 5 | 1 | 1 | Collison and Taylor discuss how AI tooling elevates high-agency, high-taste generalists. By leveraging LLM exoskeletons for coding and design, cross-functional operators with strong customer empathy can build end-to-end products without large specialized engineering armies. | |
| Twitter Board Reflections and Team Size Dynamics Under Musk | 5 | 4 | 2 | 2 | Collison asks about Musk running Twitter with an 80% reduced headcount. Taylor acknowledges that team size does not scale linearly with output, but cautions that overly austere headcount reductions can cause high-growth startups to lose market share to well-resourced competitors. | |
| Mediating the OpenAI Crisis and Fiduciary Duty to Humanity | 4 | 5 | 2 | 1 | Collison playfully asks Taylor if he attracts corporate drama after Twitter and OpenAI. Taylor clarifies his role as an agreed-upon mediator during the OpenAI board crisis and reflects on the unique fiduciary duty of stewarding non-profit AGI governance for humanity. | |
| AI Predictions for 2026: Scientific Breakthroughs and Code Generation | 4 | 4 | 1 | 1 | Taylor shares his forecast for mid-2026: AI-driven scientific and mathematical discoveries capturing public imagination, the widespread enterprise normalization of autonomous agents, and the near-total elimination of manual code writing in Silicon Valley. |
Statements from this episode (42)
Taylor: Hobbyist AI agent adoption began with a rogue open-source project
“It's probably the first, I wouldn't have predicted the first kind of broad, I don't know if consumer is exactly accurate, but maybe a hobbyist use of AI would have been this kind of semi-rogue open source project that goes through three name changes in three d…”
Collison: Mainstream AI chat interfaces still lack persistent memory in 2026
“Still in twenty-twenty-six, if you open a new Gemini chat, or if you open a new ChatGPG chat, it's basically a blank slate. There's no memory.”
Taylor: Coding agent success doesn't translate to broad knowledge work
“And so it's not like impossible for an agent to use those things, but the idea that you can like straight line from coding agents to writing the Stripe annual letter, I don't totally buy.”
Taylor: Mimicking codebases is currently the best harness for AI agents
“I wonder if in the short term, it might just be one of those idiosyncrasies of history, like mimicking a code base is actually the best way to make a general purpose agent work. And maybe over time we'll get fancier than that, but it's actually like a relative…”
Taylor: In agentic development, PRDs and intent are more durable than code
“Where there's a lot in, in the code that is more transient. You know, you might be fine to Delete that. What was the intention? What was the PRD? Like, what was the customer problem? Is actually the more durable asset.”
Taylor: Attempting to transition to a workflow where he writes no code
“I am trying to get to a world where I'm not writing code.”
Taylor: Future engineers cannot be precious about writing code by hand
“I feel like I won't be like a self,
Self-actualized software engineer in the future, if I'm too precious about that artifact, which used to be so central to me.”
Taylor: Hierarchical multi-agent architectures are elegant on whiteboards but nonsensical
“I feel like this view of a multi-agent world was you have all these agents that do tasks that are fraud detection. You know, another one over here for, you know, personalization, and then you make a super agent, it does all these things, and it looks really go…”
Taylor: Future AI agents will evolve away from Model Context Protocol
“My sense is we're making true agents over time the way we think about context and how that context is sort of like shared so that the agent that's orchestrated it actually understands sort of what's behind all these APIs and why and the history will maybe look…”
Collison: Stripe is building SSH access directly into developer accounts
“And of course, now we're building that because it's like much more relevant in the agentic world”
Collison: AI agents expose the lack of APIs for SaaS dashboard switches
“It turns out there's a lot of switches in the dashboard. Yeah. You know, there's no API for, and we are all as an industry collectively discovering that.”
Taylor: Sierra AI agents are calling each other in English over PSTN
“So, we have payers with AI agents that pick up the phone. And we have providers that have AI agents that pick up the phone and make phone calls. We have revenue cycle management companies that work to make outbound calls to do it. We've already had- ... We've …”
Taylor: Agent-to-agent phone calls are temporary until better harnesses emerge
“I think it's great that our agents have spoken over the telephone already in English. ... But I don't think it's like the long-term future because there's so much value that you can provide. And put another way, the agents using a sophisticated... Application …”
Taylor: Sierra raised SoFi's Net Promoter Score by 33 points
“We just did a great case study with SoFi and I'm really proud that we raised their net promoter score by 33 points, just because it's just so delightful.”
Taylor: Sierra reached $100M ARR in seven quarters, now around $165M
“We reached a hundred million dollars in ARR in seven quarters, a 150 in eight quarters. We're, I think, around one 65 now, one month into our you know, next quarter or so.”
Taylor: AI agents will become the primary digital interface for most businesses
“And so then you look at AI agents, and I believe most businesses, it will be their primary digital interface, and it's because it works over WhatsApp, and it works over the phone.”
Taylor: Ramp automates 90% of its customer support cases using Sierra
“Well, they're automating 90% of their cases.”
Taylor: AI automation actually increases average human support handle time
“The cases that do make its way to your customer service team can end up more complex, sort of, by definition. So what's called average handle time will actually go up.”
Taylor: Broadly accessible AI is an imperative, not a competitive advantage
“If every single company in an industry has access to technology, I would say it's an imperative, not a competitive advantage.”
Taylor: Early AI adoption offers a temporary window to shuffle market share
“This is the moment where perhaps if you have a competitive equilibrium, you can absorb this technology, use it. And you'll have this window where you can like actually like shuffle the deck.”
Taylor: Grounding AI agents is harder for famous brands than obscure ones
“It's actually easier when the internet has never heard of you and you want to make a well-grounded agent, it's actually pretty easy because there's no temptation from the LLMs to go off script. So actually, I would say. Ironically, the harder challenge is when…”
Taylor: Specialized Cantonese voice AI will be fully commoditized within three years
“So we did all this work, and in fact, you know, we, I think, have the best Cantonese support on the market. Great for us, and it's a huge selling point, and it's a technology that will certainly be commodity, commoditized in three years.”
Taylor: Applied AI startups must build ahead of models and discard code
“So the interesting part about building an applied AI company is you can't have the luxury Of waiting for all the models to catch up with your aspirations. But you know they will. So you have to have the best technology and have to be comfortable with throwing …”
Collison: Anthropic recently installed Workday
“Indeed, Anthropic just installed workday, very famously.”
Taylor: SaaS sector will likely lose value over the next ten years
“Will these companies be less valuable 10 years from now than now? I think the answer is probably yes. Will that be true for every individual company? I don't think that's true.”
Taylor: Lower marginal software costs will drive a build-over-buy enterprise shift
“One risk is that more people will build than they do now versus buy, because the marginal cost of writing software goes down. I think that'll be true for some software, particularly developer platforms and things like that that are Already being consumed in pu…”
Taylor: Database ledgers are durable against AI; systems of engagement are not
“My theory is the closer you get to literally the database is the value, i.e. A ledger, the more durable it is. The closer you get to be in a system of engagement, the less durable it is.”
Taylor: Sierra uses outcome-based pricing, charging only for fully resolved cases
“So, we do outcomes-based pricing. So, for a customer service context, that means if the AI agent resolves the case, no human intervention there's a pre-negotiated sort of rate for that, and if we do have to escalate to a person, it's free. For sales, it would …”
Taylor: AI token usage does not strongly correlate with delivered business value
“I would argue there's not a strong correlation between token usage or utilization and value. There may be, but there's not always”
Taylor: AI agents should drive long-term relationships, not isolated conversations
“I think AI agents should have memory. I think AI agents should drive relationships, not conversations.”
Taylor: Pausing model development today still leaves trillions in unrealized economic value
“I think if we paused model development, we'd still have trillions of dollars of economic value. That have yet to be realized.”
Taylor: AI adoption is bottlenecked by startups building tools over business agents
“I actually think one of the main things impeding adoption of AI is the lack of existence of all those other companies. And so many of the startups, particularly around here in San Francisco are basically doing relatively rote kind of tools around the AI rather…”
Taylor: Software engineering shifted from the most scarce to most plentiful asset
“Software engineering was the most scarce access asset in a company, and now it's the most plentiful, and I don't think we've ever lived in that world.”
Taylor: Atomic unit of AI productivity is a process, not a person
“I think the atomic unit of productivity in AI is a process, not a person.”
Taylor: Automating entire departments is science; automating specific processes is engineering
“And my point on it is, if you look at it through the lens of like an end-to-end business process, you can turn science into engineering. And I think solving legal through AI, that's a science problem.”
Taylor: Generalist product engineers using AI are worth 1,000x other engineers
“With the presence of Codex, you can produce amazing results. Those people are truly worth a thousand X other people, because it's relatively easy to find someone who's a great infrastructure engineer. Not easy easy, but like relatively. Finding someone with go…”
Collison: High-agency, non-expert engineers are ascendant at Stripe thanks to AI
“High work ethic people. Who maybe weren't the best engineers you know, previously or now, those people are massively ascendant, as far as I can tell, because they suddenly got the exoskeleton, you know, and like, they always have the ideas as to what we should…”
Taylor: Overly lean startups risk losing to slightly larger, well-staffed competitors
“If all of a sudden for some you know, clever reason you want to prove you can, the idea that, like, a competitor might have 10 people and beat you is probably more likely than even having a ten billion dollar company.”
Taylor: Altman and the OpenAI board mutually agreed on him as mediator
“Basically my understanding was I was the person that both the Existing board and Sam agreed upon to kind of help mediate the situation.”
Taylor: The OpenAI board's sole fiduciary duty is ensuring AGI benefits humanity
“The fiduciary duty is you have a duty to the mission. And that is really clarifying and interesting as well, because when you're making decisions and you realize, you know, you have, your sole duty is to ensure that artificial general intelligence benefits hum…”