Companies Keep Asking for AI. Then They Send It Back.

Hosts Sonya Huang and Pat Grady have known Elias Torres for more than a decade, across three companies. They first met him while he was rebuilding HubSpot's core platform, later backed Drift, and more recently invested in Agency, the startup he says is redefining customer experience. Torres walks on and says he has been waiting for this moment his whole life. Then he skips the script and goes straight to a hot take.

His observation is blunt: everybody is asking for AI, and when you actually deliver it, they send it back. "It's like the messiah has arrived and everybody's like no, we don't want it." The technology, he argues, is the most capable mankind has ever built, yet buyers demand a level of perfection they cannot even match when they type a basic prompt. Grady restates it as an expectation mismatch rather than a technological flaw. Torres agrees, and traces the mismatch to pre-AI software: a database lookup was a perfect string match, so people got used to narrow tools that never missed.

Now the tools are broad and imperfect, and customers split into two camps. Torres treats AI as a glass half full. If it does 10, 20, or 30 percent more than he was doing before, he is "extremely happy" because that work is gone. Other people, he says, would rather have none of it unless it is 100% perfect. He calls that a dangerous take. The example he keeps returning to is his own inbox. Agency reads his email, drafts replies, and in demos he opens the whole inbox in front of customers and hits send after a meeting. Sometimes he cringes at the wording. He sends it anyway, because he already knows he will never say anything perfectly.

Grady admits the AI product he would actually pay for is one that just takes care of email. Torres does not spend time on his inbox at all, even though that is only a side effect of Agency. The real contrast shows up in user recordings: someone will spend 20 minutes on a follow-up without changing the substance, only softening "here's my calendar link" into "let me know times." If you are working 500 opportunities, he asks, is it better to ship two perfect emails or to ping all 500? Recipients never see the polishing. They only notice whether you showed up. That, he says, is why businesses cannot scale: humans sit in the middle adding delay and speed bumps. The fix is to sort one-way doors from two-way doors, forgive a hallucinated word, and take the fear off both the sender and the receiver.

Training Data rarely dwells on origin stories, but Grady says Torres's is too unusual to skip. He grew up in Nicaragua in what he calls a communist-era, third-world scarcity. American fridges still startle him: drawers packed with snacks and boxes of food. At home, the refrigerator was empty except for a few pots, and the only groceries were whatever would be cooked for that meal. It was just-in-time living. When he asked his mother for things, she told him she would give him anything she had, and that if she was not giving it, it was because she could not. He learned a different ask: Mom, if you can buy me this, would you?

His first paid work in the United States, at 17, was a double shift: McDonald's and cleaning offices with his mother. The technology bug started earlier. After two years with a father he had not grown up with, a computer appeared in the garage. His half-brother was obsessed with cars and ignored it. Torres, in sixth and seventh grade in Los Angeles, sat down with WordPerfect and Lotus 1-2-3 and tried to make the printer work, with no internet and no user group. Back in Nicaragua his mother, a university professor and veterinarian, had a laptop that seemed impossible in that country. He became her tech support, which mostly meant breaking the machine and walking it to IT.

In 1999, at IBM, his first project was a chatbot for Blue Pages, an internal directory so 400,000 employees could look up phone numbers. He learned a lot and still felt he could not make a dent in a company that large and political. Then he met David Cancel, the first other Latino in tech he had ever seen, and chose Lookery over other startup offers because he trusted him. The team was about ten people. Torres put a desk in his bedroom in Lowell, Massachusetts, worked around the clock with three babies in diapers, and was the main engineer. The unlock was AWS: instead of begging IBM procurement for blade servers, he could press a button and get S3 and EC2. One person could orchestrate everything.

He and Cancel then founded Performable, which joined HubSpot a few years later. Grady's version, which he says Halligan and Dharmesh Shah would accept, is that early HubSpot had a great story and an okay product; Torres's team inherited the whole platform and rebuilt it. The shock was scale. Performable had about 20 customers who could text him that something was broken; he would edit code and tell them to reload. HubSpot had 5,000 customers, major churn, a product that did not work, and a team that did not function. He had to rebuild the product, the team, and customer trust, and he had to scale himself as a leader. The lesson he ranks highest is hiring.

The roster still makes him proud. He hired Andrew Bialecki before Klaviyo, Christopher O'Donnell before HubSpot CRM, plus Whitney and Jared, who as an EVP still runs HubSpot's systems. It is not obvious that a mid-stage marketing company in Boston would attract people like that. Torres's theory is that many people are exceptional if you give them a shot and show them what an outlier looks like. He is not sure how much of his own success is the kid in the garage versus the people, including Halligan and Shah, who later showed him what was possible.

He does not hire for credentials, for where someone came from, or for what they claim they already did. If a candidate arrives with all of that, he rules them out immediately. Grady supplies the word "suspicious." Torres goes further: "It's just sus. Totally sus." What he tests in the room is whether they can keep up in conversation, whether they have hunger, and whether they have grit. Then he brings them along for the ride and lets them choose the path after that. Several of those people were doing the crappiest jobs imaginable when he found them, which is the point. He was tinkering with WordPerfect in a garage. They were in the same kind of place.

Drift, his second company with Cancel, was a reaction to HubSpot's time-to-value problem. Inbound marketing asks customers to write blogs, rank in search, collect leads, and become a brand, a long journey at a low price that Torres still calls an unbelievable achievement. He and Cancel kept asking how to make value show up the same day. They were patient in a painful way. They spent about two years wandering, first on a Glassdoor-like anonymous workplace app inspired by Secret, then pivoting almost monthly. Every time Grady flew to Boston, the answer had changed: a new Dropbox, a new Pinterest, a new SurveyMonkey. Torres would tell the team the last idea was not good and they would try another. The lesson he took from the desert is that experience compounds, and you have to go back to what you know.

They landed next door to HubSpot without becoming another marketing suite or another CRM. The product that finally took off was a chatbot receptionist on the homepage. Instead of a form, it asked questions, delivered some value, booked a calendar, routed the lead to the right rep, and pinged that rep's phone. Deploy it and leads start arriving without extra human effort. It worked, and it was still shallow. It only handled the first connection. Customers came back and asked why the bot did not recognize them, pick up the old thread, or know about tickets they had already filed. Torres had a dream of always knowing everything about the customer and applying it in every exchange. The technology to deliver that did not exist yet.

The second failure mode was the growth playbook itself. They built product and brand fast, then defaulted to hiring more bodies and raising more money, assuming spend would sort itself out later. Torres now calls that instinct fundamentally flawed, and he is not sure the industry has absorbed the point. In the age of AI, he says, proving it wrong is the whole point. He is an extrovert who liked customers texting and calling him. Once Drift had thousands of customers, he could no longer know them. When a customer-success manager called about churn, it was usually too late. He puts his save rate at about 20%, and only if he promised the world, which felt hypocritical.

Headcount made the distance worse. He keeps saying 600, then 800 people. His calendar filled with promotions, onboarding, recruiting, entertainment, and town halls. That work, he argues, took their eyes off the customer and the product, and that is where they paid the price for not building the legendary company they had dreamed about at the start. Drift was still worth more than a billion dollars. Grady calls him a walking American-dream story and asks the obvious question: why do it again?

He did ride into the sunset, for about seven months. Travel felt empty. The trip that actually disconnected him was kite surfing in Tatajuba, a remote stretch of northeastern Brazil. Mornings were spent staring at empty beaches, telling himself he should learn to meditate; afternoons were on the water. Then ChatGPT launched. He showed it to everyone around him and could not explain how it knew the answers. The first day he was truly off work, he says, was the greatest moment in mankind, and he decided he had to go back in.

For years he had filed himself as an impostor: the wrong schools, not in Silicon Valley, in Boston, English learned later, a thin network, not enough investor relationships. After Drift he could no longer make that case. He had built systems, served millions of users, made revenue, and knew the people. He was not going to sit on a beach while people who "don't know anything" started AI companies. There is a more earnest reason, which he admits is cheesy. As a Latino he feels an obligation. Most people would die to be in his position. Going back to that empty fridge in Nicaragua, squandering earned time and opportunity on kite surfing is a waste. People want to see what someone who looks like him can achieve. He told himself he could not sit this one out.

He called Sequoia, showed up at AI Ascent in early 2023, and got reconnected. OpenAI, then around 200 people, could not talk to every enterprise that wanted to put the model in front of customers, a problem that looked a lot like Drift. A cohort of founders started with AI services; Torres did too, working with companies including the NBA, Red Bull, and Klaviyo. Andrew Bialecki's ask was to make every customer-success manager a world expert in Klaviyo and ecommerce marketing on day one. Torres's first answer was no. If he actually did that, he joked, his company would be bigger than Bialecki's. Bialecki told him to just do it. Roles had reversed: the person Torres had once hired was now pushing him.

They built a model that understood every customer and every interaction, classified usage, and delivered tailored advice. Then the rest clicked: what if the quarterly business review wrote itself, with a video that explained it? That same day Brian Halligan texted. He had been romancing Torres for a while, and now he had an idea: disrupt the entire industry, because people cannot scale. They called Sequoia, and Agency was born. The problem, as Torres defines it, is that the world still delivers customer experience with humans, and human-led experience is almost impossible. Even the pampered version is a joke. In a hotel room that costs thousands of dollars a night, asking for Tylenol can draw a $4 delivery fee, and food sent upstairs another $7. That, he says, is as good as customer experience gets today.

Pressed for traction, Torres says more than 50 customers are using the product and that he does not care about the vanity numbers. He was trying to add more logos when Halligan told him to stay with the ones he has and make them wildly successful. He agreed. The harder confession is about large language model costs. He does not know how they survive them, and he calls the economics "an entire pyramid scheme." What he is sure about is the commercial posture. In the past, the minute something loaded in a browser he wanted to charge a dollar. This time he does not want the money yet. He wants to know customers are obsessed, that the product is sticky, and that Agency is doing everything at the company.

He has watched too many founders confuse revenue with product-market fit. "People have $50 million in revenue and they're delusional that that product works." Most SaaS he ever built, including his own, created more work for the user. This time he wants something that takes work off. Patience may hurt him or it may not; he credits investors like Grady with not rushing him. The goal he said out loud early, when people told him he was crazy, is a company at a billion dollars of revenue with fewer than 100 employees. Most of the team is engineers. In a typical software company, headcount and spend pile into go-to-market and customer experience. He wants to prove that layer can be eaten by AI, so the companies of the future can actually get built.

The product map is a HubSpot habit Grady recalls as "when they zig, we zag." When Torres started Agency, the zig was a swarm of AI SDR companies going after the front of the funnel. His zag was to let them have the front end and take the back end of the business, where the pain and the value actually sit. A year later Grady thinks that looks smart. Torres is less impressed by the rest of the market. In the most creative country in the world, he says, everyone is shipping the same chat app, meeting recorder, note summarizer, and email drafter.

Asked what AI can and cannot do today, Torres separates the fullness of time from the present. In theory AI is infinitely capable. Today it is "not that smart." You write paragraphs to maybe get the right answer, 42, and if you change a word or try again you get a different one. Human brains, he says, cannot cope with that randomness. The product question follows: are you trying to find a better sentence to send the customer, or are you letting AI actually run the experience? If you cannot even trust a draft you are about to send yourself, how much worse is it to let the model decide for 10,000 customers a day on its own?

Agency's first layer looks like everyone else's: a customer 360 that gathers information in one place so you can ask questions. The next step is where most tools die. "If you require a human to go type into that chat box, you're dead on arrival." He is done building software that merely helps a person do a task. Typing into ChatGPT bores him. He is a spoiled CEO now and wants to tell the system what to do once and have it do it 10,000 times. Agents, as a buzzword, get no sympathy. A meeting summary in his inbox is worthless. He can always ask that question. The useful question is what happens next.

Customers still request dashboards. He keeps asking "and then what?" They want to click through, chase renewals, tell a rep, have the rep email the customer. His answer is: why can't the system just email the customer? That is the deprogramming job, and it is the hardest part, because the business world has operated the old way for too long. At Halligan's birthday party a retired VC delivered the line like an oracle: forget everything you have learned, forget pattern recognition, forget experience, those things lock you. Experienced enterprise founders, Torres took it, are their own biggest enemy. The old map is the thing to unlearn.

Getting AI to read a dashboard, make a concrete decision, and do it the same way every time will take a long time and a lot of prompting. Internal inefficiencies are worth fixing. The valuable work is still delivering something directly to the customer. Asked what he is most optimistic about, Torres gives the same word as the company name: agency. People have spent history on hard work. He wants them to stop grinding and do what humans are actually best at, taking initiative, choosing a purpose, choosing a customer, choosing an audience, and going to serve them. AI cannot take that away. Used well, it only amplifies it.