Bret Taylor

the story of bret taylor is, in a way, the story of the last twenty years of consumer software, told through one person who kept showing up at the exact moment the ground was shifting. he was born in 1980. he studied computer science at stanford. and then, in his mid-twenties, he landed at google at the precise moment google was figuring out what it was going to be beyond search.

the thing he's best known for from those years is google maps. taylor was on the small team that built it, and he's widely credited as one of its creators. it's worth sitting with what maps actually was as a piece of engineering, because it tells you something about how he thinks. at the time, web pages were static. you clicked, the whole page reloaded, you waited. maps did something that felt like a magic trick — you grabbed the map with your mouse and dragged it, and the world slid under your cursor, and new tiles streamed in without the page ever reloading. that pattern, fetching data in the background and updating the page in place, is the thing that later got a name — ajax — and it basically defined the web app era. taylor was standing right at that hinge.

he also, around this time, met a colleague named clay bavor. taylor hired bavor as an associate product manager at google in 2005. hold onto that name, because it comes back twenty years later.

then taylor did the thing that becomes a pattern with him: he left the big company to build a small one. he co-founded a startup called friendfeed, an early social aggregator — a place to pull together everything your friends were posting across the web into a single feed. it was clever, it was early, and in 2009 facebook bought it. and this is where taylor's career takes its next turn, because he didn't just get acquired and drift off. he rose. by 2010 he was facebook's chief technology officer. the ideas friendfeed had been playing with — the real-time feed, the "like" button — flowed into the core of facebook itself. and then, true to form, in 2012 he left again to start another company.

that company was quip. quip was a bet that documents and spreadsheets should be collaborative and mobile-first, built for a world of phones and teams, not desktop files emailed around as attachments. it was a good product fighting a very hard fight against microsoft and google. in 2016, salesforce bought quip for around seven hundred and fifty million dollars, and taylor went inside salesforce. and once again, he climbed. he became president, he became chief product officer, and eventually he became co-ceo of salesforce, running the company alongside its founder marc benioff.

around the same stretch, he took on a role that would put his name in the middle of one of the messiest corporate dramas of the decade. he was chairman of twitter's board. and he was chairman during elon musk's takeover of twitter — the on-again, off-again, will-he-won't-he acquisition that ended up in delaware court before it closed. taylor was the person on the other side of the table, representing shareholders, holding musk to the deal he'd signed. remember that too, because taylor and musk end up circling each other again.

so that's the shape of the career before this moment: maps, friendfeed, facebook, quip, salesforce, twitter's board. a person who alternates between building inside giants and starting things from zero, and who keeps ending up in the room where the important decisions get made.

which brings us to why he's worth an entire episode right now. because taylor is currently sitting at the intersection of two of the biggest stories in technology at the same time. he is the chairman of openai's board. and he is the co-founder and ceo of an ai company called sierra that, this month, raised almost a billion dollars.

let's take the money first, because the number is genuinely arresting. sierra announced a nine hundred and fifty million dollar funding round — a series e — led by tiger global and google's venture arm, gv, with existing investors like benchmark, sequoia, and greenoaks also participating. that round values the company at fifteen point eight billion dollars. and it means sierra now has more than a billion dollars of cash on hand.

here's what makes that number worth pausing on. sierra is about three years old. taylor co-founded it in 2023 — with clay bavor, the same person he hired at google back in 2005 — and it launched its product in early 2024. now walk the funding history, because the slope of it is the story. a hundred and ten million dollars in early 2024, led by sequoia and benchmark, at a valuation of around a billion. a hundred and seventy-five million in the fall of 2024, at four and a half billion. three hundred and fifty million in september of 2025, led by greenoaks, at ten billion. and now nine hundred and fifty million at nearly sixteen billion. the valuation roughly doubled, then more than doubled, then went up again, all inside about eighteen months.

and underneath the valuation, there's real revenue moving fast. sierra crossed a hundred million dollars in annual recurring revenue in november of 2025 — that's less than two years after launch. by early february of 2026 it was reported at a hundred and fifty million. and by this spring, estimates put it around two hundred million in arr, up from roughly twenty-six million at the end of 2024. taylor described that growth timeline as unprecedented in traditional software, a sign of what he called intense demand in the market. the company says more than forty percent of the fortune 50 are customers — names like weightwatchers, sonos, siriusxm, adt, ramp, rivian, rocket mortgage, nubank.

now, a healthy dose of skepticism is warranted here, and taylor himself has supplied some of it. even at a hundred million in arr, the company was carrying something like a hundred-times revenue multiple. that is an enormous number. and this is the same bret taylor who, in january, told cnbc that ai is "probably" a bubble and that he expects a correction in the coming years. so you have a founder raising at a sixteen-billion-dollar valuation while openly saying the sector he's in is probably inflated. his framing on the raise was competitive and blunt — that there's a lot of competition, that sierra is, in his words, multiples larger than the next biggest player, and that they're investing aggressively to expand their lead. in other words: the market is frothy, and the way you survive froth is to get big enough that you're still standing when it drains.

and then there's the other hat. taylor is the chairman of openai's board. he got that job in the most dramatic way possible. in november of 2023, openai's board abruptly fired sam altman, the company nearly came apart over a weekend, and when altman was brought back, it was under a new board — with bret taylor as chair, alongside larry summers and adam d'angelo. taylor was, essentially, the adult brought in to stabilize the most important and most volatile company in ai. he's held that chair through everything that's come since, including openai's restructuring and its escalating fight with elon musk.

that fight reached a climax just weeks ago. musk — who co-founded openai back in 2015, put in something like thirty-eight million dollars, and then resigned from the board in 2018 — sued openai and altman, arguing they had betrayed the nonprofit mission and turned a charity into a for-profit engine for personal enrichment. he sought damages reported at a hundred and fifty billion dollars. it went to a jury in oakland. and in may, that nine-person jury handed altman a decisive win. but notice how they did it: they didn't actually rule on whether openai betrayed its mission. they found that musk had waited too long — that the statute of limitations had run out before he filed in 2024. they deliberated for less than two hours. the judge, yvonne gonzalez rogers, dismissed the case. musk called it a technicality and vowed to appeal. so the central question — what openai owes to the mission it was founded on — is still, technically, unanswered. and taylor is the chairman sitting on top of that unresolved question.

he's also become one of the more measured public voices on ai regulation, which is part of why he keeps showing up in the news. his general posture, as of this summer, is that he's heartened that governments and companies are taking ai regulation seriously — a deliberately calm, institution-builder's tone, which is very on-brand for a guy whose whole career has been about being the stabilizing presence in the room.

okay. that's the who and the what. now let's go under the hood, because the technology sierra is building is more interesting than "chatbot for customer service," and the details are where the real lesson lives for anyone building with this stuff.

start with the word "agent," because it's doing a lot of work and it's badly overused. when taylor says agent, he means something specific. a chatbot answers a question. an agent takes an action. the distinction is everything. a chatbot for a bank can tell you your balance is low. an agent can move money, dispute a charge, close the account. sierra's whole thesis is that the valuable part isn't the conversation — it's the ability to actually do the thing at the end of the conversation. their agents don't just retrieve an answer; they process refunds, open and update tickets, change subscriptions, and — this is the part that's technically hard — they collect card and ach payments. sierra built what it describes as the first level-1 pci-compliant conversational ai platform, which in plain terms means an ai agent that's certified to handle credit card transactions end to end inside a single conversation. that is a compliance and security mountain that has nothing to do with how clever your language model is, and everything to do with why enterprises actually buy.

so how is the thing built? sierra's platform is called agent os — they're now on agent os 2.0 — and it's not one model, it's a stack of products around the model. let me walk the important pieces, because the architecture is the argument.

first, retrieval. the agents use rag — retrieval-augmented generation — to pull answers from the specific company's knowledge. this matters because the base language model doesn't know your return policy, your billing rules, or what "gold tier" means at your company. rag is the mechanism that grounds the model's response in your actual documents and data at the moment of the conversation, rather than relying on what the model happened to memorize during training. it's the difference between an agent that sounds confident and an agent that's correct about your specific policies.

second, actions and integration. an agent that can only talk is a toy. to be useful it has to reach into the customer's real systems — the crm where the customer record lives, the billing system, the erp. sierra has an integration layer for exactly this, so the agent can read state and write state in the systems of record. this is, honestly, where most of the hard engineering in enterprise ai actually is. the model is a commodity you can rent. the plumbing into a twenty-year-old billing system, with the right permissions and the right audit trail, is not.

third, memory. sierra has something called the agent data platform, which gives agents persistent memory and context across interactions. so the agent isn't starting from zero every time you call. it remembers the last conversation, the open issue, the context. anyone who's built with language models knows that context is both the whole game and the hardest thing to manage — the model itself is stateless, so all of the state, all of the memory, has to be engineered around it.

fourth, control. and this is the part that separates a demo from a production system you'd trust with millions of customers. language models are probabilistic. they can be confidently wrong. they can be talked into things. if you're an airline or a bank, an agent that "hallucinates" a refund policy isn't a cute bug, it's a financial and legal liability. so sierra's platform is really an apparatus for constraining the model — guardrails, defined workflows, and human escalation. they have a product called live assist for handing a conversation off to a human when the agent hits its limits. and agent studio 2.0, their configuration environment, introduced structured concepts they call journeys and workspaces — essentially ways to define the allowed paths an agent can take, so that non-engineers on a business team can build and manage these workflows through a low-code interface rather than everything going through bespoke engineering. that's a tell about where they think the bottleneck is: not the intelligence, but the configuration and the safety.

then there's the piece that shows you where all of this is heading. it's called ghostwriter, and it is an agent that builds agents. you feed it your standard operating procedures, your support transcripts, your process documents — even photos or audio recordings — and it generates a production-ready agent across voice, chat, and email, in more than thirty languages, with guardrails attached. sit with that for a second. the labor of building the agent — reading the manuals, mapping the workflows, writing the logic — is itself being handed to a model. that's the recursion at the center of this whole industry right now: using ai to compress the cost of deploying ai.

one more technical shift that's easy to miss but is, i think, the most important one. voice. sierra launched a voice product, and by october of 2025, voice had already overtaken text as the primary channel for its agents — on pace to power hundreds of millions of ai phone calls. that's a big deal, because voice is brutally harder than chat. in text, you have time. in a phone call, latency is the enemy. a human expects a response in a few hundred milliseconds, and every pause where the model is "thinking" feels like the line went dead. so a real-time voice agent has to do speech-to-text, retrieval, reasoning, action, and text-to-speech, and it has to do all of it fast enough to feel like a conversation, while handling interruptions when the caller talks over it. the fact that enterprises shifted call-center volume onto this so quickly tells you the latency problem got solved well enough to be real. and voice is where the actual volume is — something like eighty percent of customer service still happens over the phone.

now pull all of that together and you can see what taylor is actually claiming. his thesis, which he's laid out at industry conferences, is that most enterprise software goes largely unused — that companies buy these big complicated tools and employees barely touch a fraction of the features. and his bet is that the future isn't a better interface for humans to navigate; it's an agent that navigates on your behalf, so the human never has to learn the software at all. "agents eating software" is the slogan. and notice how sierra's business model encodes that belief. they don't primarily sell seats or subscriptions. they charge on usage and, increasingly, on outcomes — pay per resolved conversation. that pricing is a philosophical statement. a per-seat license assumes a human sitting at a screen. pay-per-resolution assumes the work is being done by the system, and you pay when it works. if he's right about agents replacing interfaces, the whole per-seat model of enterprise software — which is basically how salesforce, the company he used to run, makes its money — has a problem.

so what does a cto actually take from all of this? let me try to be concrete, because this is the part that matters more than the valuation gossip.

first, the pricing signal is the loudest signal. watch outcome-based pricing very carefully, because it's a wedge aimed straight at the seat-based software you already pay for. if your vendors start charging per resolution instead of per user, your budgeting model changes. the good news is that cost scales with value delivered. the bad news is it gets much harder to forecast, and you can lose the predictability that made saas easy to plan around. and there's a real second-order effect that's already showing up: teams are blowing through ai budgets faster than expected once they open the door to agents, because usage-based spend has no natural ceiling the way a fixed seat count does. so if you're adopting agentic tools, put metering and cost controls in on day one. don't find out at the end of the quarter.

second, and this is the durable lesson from sierra's architecture: the moat is not the model. every serious player is renting or building comparable models, and the frontier keeps moving. the defensible work — the stuff that took sierra a real engineering org to build — is the boring, hard layer around the model. the integrations into your systems of record. the permissions and the audit trails. pci compliance. persistent memory. guardrails and the escalation path to a human. when you evaluate an ai agent vendor, or when you decide whether to build in-house, that layer is your real diligence checklist. ask how it handles the case where the model is confidently wrong. ask what it's certified to touch. ask how it fails safe. a slick demo tells you almost nothing; the failure modes tell you everything.

third, think about where the human goes. taylor's public position on jobs is more careful than the doom-and-hype poles of this conversation, and the product itself reflects a specific answer: live assist, the human-escalation tooling, is a core part of the platform, not an afterthought. the near-term reality this points to isn't "no humans." it's a smaller number of humans handling the hard, ambiguous, high-stakes cases, sitting on top of a large volume of automated resolution. if you run a support org, or any ops function that's mostly structured workflows, that's the shape to plan for — and it changes who you hire. you need fewer people doing tier-one repetition and more people who can design workflows, review agent behavior, and own the exceptions. the job of "configuring and supervising the agent" is a real job now, and ghostwriter is a signal that even that job is going to get compressed.

fourth, hold the macro picture honestly, the way taylor himself does. the same man raising a billion dollars is telling you this is probably a bubble and to expect a correction. those two things aren't a contradiction — they're a strategy. the lesson for a cto isn't "wait for it to pop" and it isn't "bet the company on the frothiest vendor." it's that consolidation is coming. some of the two dozen agent startups in your inbox will not exist in three years. so when you pick a partner for something as sensitive as customer transactions, durability is a feature. capitalization, real revenue, and a plausible path to standing after a correction should weigh as heavily as the quality of the demo. that's an uncomfortable thing to admit, because it favors the incumbents-in-waiting, but it's true.

and finally, the thing that ties the two taylor stories together. the man building the agent company is also the chairman steering the company that makes the models those agents increasingly run on. that vertical closeness between the model layer and the application layer is one of the defining structural questions of this era, and the musk trial — for all its billionaire theater — was really a fight about who controls that stack and on whose terms. as a cto, you are building on top of that unsettled ground. the governance of these companies, the regulation taylor says he's heartened to see taking shape, the ownership of the model layer — none of it is resolved. it's not an abstraction. it determines your pricing, your dependencies, and your risk. the person who spent his career being the stabilizing adult in the room is now sitting at the center of the least stable, most consequential platform shift we've had since the one he helped kick off, twenty years ago, by letting you drag a map with your mouse.

that's the through-line worth remembering. the tools change completely. the pattern — build the boring hard layer, own the moment the ground shifts, and price for the world you think is coming rather than the one that exists — that part, bret taylor has run four or five times now. it's worth watching whether it holds one more time.