Elias Torres sold a billion-dollar company and called it a failure. Agency is the redo.

Sarah Guo introduces Elias Torres as a repeat founder and CEO of Agency, a company that wants every firm to make its customers successful. He has founded four companies before this one, led engineering at HubSpot, and most recently sold Drift to Vista Equity Partners for more than a billion dollars. The fifth company is the one he actually wants to talk about. The path there is the part he finds interesting: Nicaragua, first generation, no English, the back of a McDonald’s reading printouts, then three companies with Sequoia, IPOs, and Boston rather than the Valley.

He spent ten years at IBM and left because he could not stay inside it. He needed to be free and to have a chance at huge impact. In 1991 or 1992 he was pulled into national math contests in Nicaragua, placed third in the country, and went to Mexico to compete against kids from across Latin America — a mind-boggling first look at how far the ladder went. In a low-income Tampa public school, still without English, a teacher asked him into Mu Alpha Theta. Word problems wrecked him. The other problems did not. Math, Guo notes, has been minting future tech people for a long time.

Brian Halligan bought Torres’s company Performable and told him HubSpot would go public in a year or two. The firm had $30 million in revenue and 200 employees. Torres had no idea what going public meant. Halligan put him and David Cancel in charge of product and engineering. Three years later HubSpot listed as a billion-dollar company; revenue had gone from $30 million to $130 million. The naive thought that followed was the seed of Drift: if Brian can do this, why can’t I? Zero to $100 million should be easy.

It was not easy, but it did work. Drift took eight years of hard work to $100 million in revenue and 800 people, and the sale was more than a billion dollars in cash. Torres still talks about it as a failure, which Guo flags as an uncommon view of that outcome. He is happy with his life — he had almost no food growing up, and every year has been better than the one before. The wound is relative. He left HubSpot at a billion dollars and watched it grow to $30 billion. Building Drift, the goalposts moved. Signing night was 11 p.m., in bed, tired, a lawyer’s text that the papers were done. He felt nothing. The next day felt weirder. He had lost the dream of building a $30 billion company or taking one public. He had worked straight for 45 years, barely taking vacations, and if you had asked him then whether he would start something else, he would have told you not to. Starting a company, he was saying, was stupid.

The energy came back in November 2022, which Guo cannot resist pairing with Conviction’s October launch — he stole her thunder, she jokes; he says she stole his. He was in Brazil on the first real vacation after the acquisition. He had stayed at Drift almost two years. He kitesurfs now. A poor immigrant, he says, does not have hobbies; he has them for the first time. Four friends, morning session, afternoon session, ocean breeze. In the middle of that he tried to explain ChatGPT to them and stalled. He knew BERT, Word2Vec, Transformers. He did not know how to explain why this was different. The first feeling was that he had missed the boat again: why wasn’t he at OpenAI, is the game over?

He had already concluded that the only thing he knows how to do is build, and he wanted to learn. A networking intro to OpenAI turned into a plea: we are drowning. Two to three hundred people, customers begging for help implementing their own solutions, big firms tried, no bandwidth. He said he wanted to do that. He had never used the GPT API. They looked him up and asked if he was sure. He said yes. He started at the bottom, support tickets and explanations, people asking who he was. Then the work got interesting: NBA, Ticketmaster, Red Bull, a trip to Dubai, Klaviyo — whose CEO, Andrew Bialecki, is an old Performable colleague — and a small team implementing agents and LLM apps. For an engineer, he says, it was the dream: building solutions this intelligent.

Agency is the leftover lesson from four companies. Startups know how to raise money, hire, build product, sometimes sell and market. Torres does not believe they know how to take care of customers, especially in B2B, especially at tens or hundreds of thousands of accounts. That problem is about ten years old and companies are still struggling. He landed on it because of Klaviyo. Bialecki needed to scale customer success across hundreds of thousands of customers. Torres had already built an AI SDR inside Drift in 2019, three years ahead of the current wave. Applying marketing experience plus LLMs — benchmarking, reporting, insight generation, idea creation at a level of detail humans cannot sustain — he realized he could do this for every B2B company.

Five-year CS, in his mouth, is not a hotter software category. He thinks CS is a misnomer, which is where people get lost. VCs count the companies in the bucket, add up TAM and valuations, and decide if they care. He is not necessarily interested in CS. He is interested in customers. He has lived in the post-sales world for about 15 years and calls himself a tried-and-true customer-obsessed founder: in every company he is the one who services the accounts. At HubSpot the book went from 20 customers to 5,000. People called and asked him to get on the screen and fix HubSpot for them. He had to say no — he had 5,000 customers. Every time he stopped serving someone as if they were the only one, his soul got crushed. That is the thing he wants to solve: not a CS motion, but giving power and agency back to the customer.

The analog he wants in software is ordinary local service. A man down the street stores his car in winter, keeps a space, texts, picks him up, stays flexible; Torres is in charge of the relationship. His barber is Mel. He booked two Saturday haircuts for a son; the second son asked to come; Torres told the kid to text Mel and see if he could squeeze them in. Mel did. Can you have that with a ServiceNow today? He wants businesses to put the customer in charge, at scale, with hundreds of thousands of accounts. Guo recognizes the instinct from venture: she does not want four layers of talented people between her and founders, because they lack context and care less, even if that means the firm does not scale in the usual way.

Torres wants the same break with startup orthodoxy. At Agency there will be one email address, elias@agency.inc. He admits it is dangerous; everyone will write. He still intends to take care of all the customers. At Drift, 800 people, about 250 in the customer organization, did not solve it. That headcount cannot happen again. Every other company is struggling with the same ratio. The only way he sees to keep the relationship is AI.

He does not want 800 people this time. The number he repeats is a billion dollars in revenue — value, he says, is too easy — with 100 people. How do you not hire the other 700? Question everything. He is a fan of Elon Musk’s rocket-part logic: do not take marketplace prices as given; ask what it costs to make the part. You make a billion either with something lots of small businesses can acquire for free at tiny marketing cost, or with something very expensive for enterprises. He is picking enormous problems for big companies.

Then he goes after Paul Graham. “Do things that don’t scale” is game over. Now you only have to do things that scale. He already knows the old move: throw a body at it, it works for a year, it breaks. Guo had tweeted about hiring an intern; someone — he thinks Pav — replied that you could throw Devin at it. That is the thinking he wants. Do not do the unscalable thing. Solve it from the first principles: customer intelligence, communication, a deep grasp of the problem, better pricing, the right relationships with the right customers. Agency is only a few months old and he already feels further ahead than Drift was at almost three years — on the problem, the market, positioning, team, engineering, infrastructure, branding, marketing. Partly AI. Partly 30 years of scar tissue. It feels natural.

Guo asks what makes someone worth being in the last hundred. He has stopped caring enough to be polite. This is his first time in incubators — three of them, fifth company — and watching first-time founders was a refresher on mistakes he has made too many times. Someone says the right things, they are available, they will get you to a billion 25% faster. He no longer buys it. He has hired every C-level multiple times, from the best companies, including people who went from zero to $200 million in four years. Nobody joins Agency until they have worked for him as a contractor. He will be as real as he is with the current team. If you do not deliver at a world-class level in one or two weeks, there is no room. He is focused mostly on engineers. Everyone needs a clear role. It is very difficult to be part of this 100.

The hardest thing about building Agency, he says, is always product, and always change management. The lazy story is that you put an LLM on it, generate some stuff, summarize a meeting, and boom, here is the next CRM. LLMs solve one tiny part of the problem. You still have to build the product that solves a specific solution. Large post-sales organizations are chaos. Someone has to go in, understand them, talk to them, nurture them, see what is broken, and slowly transition. He has talked to CEOs of public companies. None of them say throw it all away and swap it tomorrow.

Trust is lopsided. Nobody cares if an LLM summarizes a meeting; they can read faster, great. Ask a customer-success manager to trust that output in an email to a million-dollar customer and they question every word. When they write their own emails they do not question their words, and those emails are just as bad. There is a long road, and it is not a problem one foundation model fixes. He is not spending the hour on whether AI will solve everything. He is listening to customers and asking whether they will come on the journey to where he wants to take them.

After 30 years of building software, Torres is dismissive of most of it. People worry AI will enslave humanity; he tells them they are enslaved today. Who wants to be the sales rep who finishes a meeting and then enters a task to remind themselves to remind the customer? The emperor has no clothes and nobody wants to say it. Uber is the exception he will praise: you need to go somewhere, you call, it shows up, it takes you there. Why enter information into systems that do nothing for you? Should Agency install Salesforce? He calls it a monstrosity you hire someone to configure, then cannot use. The configurator does not know your business or your customers, so they hand you an antiquated way of working. Then you hire people to type into it, then people to monitor those people. At what point did you talk to the customer? He tells CEOs to throw their Salesforce instance in the trash. They are busy configuring it and sending it data they never use.

The replacement is software that acts without being asked. His calendar has a JetBlue flight because he is speaking in San Francisco. It should read that, find everyone who always wants coffee there, message them that he is in town on these dates, lock two-hour slots, name the coffee shop, and put the customers he is chasing at the front of the line with notes two weeks out, repeating every three days. When you can text someone and it just gets done, that is wealth. He does not feel wealthy from his software today. Old-world infrastructure — places to store information, the internet — still has a job. The old wrappers of databases at the solution layer die unless they adapt quickly. New CRMs that keep the same views and tables plus an AI-computed column do not count. He wants software that is almost invisible: do not bother people, do not announce what you will do, just do it, get verification, learn preferences, personalize. Once there are examples, and LLMs let more people build faster, the big rebuilding begins.

Guo says the rant lands. She is talented and very disorganized, and she has felt bad about it for most of two decades. Maybe the question is why software makes her file, label, and jam life into a project-management database. Software should accept humans as they are. Torres’s advice is the bucket theory: skills sit in levels from mediocre to terrible to great, and it takes something like ten years to move a bucket. Focus on the skills where you only have to climb one level. Do not drag your poorest skill to the top. Let it be. AI should handle organization: tell her what to ask, prepare context, do not make her lift a finger. Interviewing people is the part a model cannot do because it is her. His own English writing was terrible for decades; before ChatGPT he wondered whether to learn it. Now he hands it to the model. They joke that next time the avatar can take the prep and they can just get a beer. The relationship, he says, is what matters. They one-shotted the podcast.