The Man Who Took HubSpot CRM to $40 Million Says the Data Model Has to Be Thrown Out

Pat Grady recorded this episode of Sequoia’s *Training Data* on the road in Boston with Christopher O’Donnell, founder of Day.ai and the person who built HubSpot’s CRM. Grady’s frame is blunt. Enterprise software has three super-categories — CRM, ERP, and productivity — and Salesforce is the most valuable company to come out of the cloud transition, dominant in CRM for two decades with one exception. HubSpot mounted the successful assault: the most beloved CRM in the market, more than a billion in revenue, growing, with real margins. O’Donnell is the operator behind that product.

He joined HubSpot through an acquisition in 2011. In April 2023, five or six months after ChatGPT’s moment, he and his old partner Mike Pici started Day.ai to do CRM again, this time AI-native. At HubSpot they had snuck in through sales-acceleration plugins before taking the system of record. At Day they are not popping out of a cake. They are starting with the CRM. His judgment is hard: the existing data model cannot carry what AI makes possible, so it has to be rewritten from scratch.

Ask a hundred CRM users the classic interview question that gets to core pain, he says, and a hundred of them give you some flavor of the same sentence: they are scared of things falling through the cracks. That is the product problem Day is actually solving, and it is also the human one. People come to work to add value. They also want to feel respected, worthy of the paycheck, not about to be fired. A system that remembers is a system that lets them operate at that level.

O’Donnell tells his HubSpot years as three chapters. The first was marketing email, right after the famous Owner Ollie versus Marketing Mary ICP decision that business schools still teach. The company needed an email system. He and Whitney Sorenson, still HubSpot’s CTO, rewrote and re-platformed from zero: four months to get off a well-known enterprise email app and onto something they used internally, seven months to market. O’Donnell built the front ends because there was nobody else; Sorenson did the backends. They took the company public and, he argues, won the category. Marketo went public, then private, and faded. The whole time they had an eye on a universal contact database, one stack that would grow past email.

The second chapter started at the end of 2013. He stood up a product-management practice, handed marketing to a team “like fishermen throwing the fish back in the water,” and started over. The founders approved a startup within a startup. They hung a pirate flag, sat on their own stack, billed through their own Stripe, and began as O’Donnell and Pici. Wallet share was leaking to CRM plugins — enrichment, better decks, better emails — so they started with the simplest possible thing, email open notifications, and grew it into a sales-acceleration suite. A year or two in, the company decided to do core CRM as well. Product and engineering were already set up for it. The team grew to about a hundred people and from zero to about $40 million in revenue, then was acquired back into the mothership department by department: engineering first, then product, then sales, customer support last. He describes it as splitting one acquisition into four, which made it more likely to stick.

Part of the mandate was to reinvent go-to-market from the bottom up. There was no PLG blog yet. They were inventing demand generation that rotated into sales alongside companies like Expensify and Atlassian, with no textbook. The breakthrough he still remembers by where he was sitting: Pici and Mark Roberge walked in with a spreadsheet of lifetime value by product area, once people had hit their usage limits. The third chapter was four or five years as steward of the overall product team. Stanford is writing a case on the skunkworks now. In hindsight it looks smooth. In the room it was messy, ripped from Clayton Christensen: every extra dollar in the core product still yielded margin, and nobody knew how Brian Halligan’s second S-curve was supposed to grow except by walking it.

Why does the world need an AI-native CRM? O’Donnell counts three problems that existing systems are never going to shake. First, data. Even the best CRM implementation on the planet probably has 40 or 50 percent of the data it *thinks* it should have, let alone what is now possible to capture. Erik Munson, Day’s founding engineer, talks about compression. Legacy CRM is a photograph down-sampled into pixel art, a few blocks, like a crypto ape and maybe not as valuable. People have to type the data in, so you only add a field if you believe someone will actually fill it out. Otherwise it stays empty. That is the governing principle of the old model.

Second, workflow. Watch a heavy CRM user and you will see forty tabs of the same product across the browser. It is easy to lose your place, easy to lose the plot, and you are not working back and forth with the system to understand a relationship. Third, work product. Users do not sit down, ask questions, and walk away with an internal memo, a knowledge-base article, an email draft, or meeting prep. That happens in Notion or Google Docs. O’Donnell finds this absurd. The CRM already holds the full history between the company and its customers. A sales manager should sit down, ask what to watch this week and where to help, and leave with a list.

Incumbents, he thinks, will do a decent job of going from 8-bit to 16-bit. The AI-native opening is PlayStation 5 with ray tracing — Super Mario Bros. to Elden Ring. Show his parents a modern game and they think it is footage of the real world. To support that, you cannot store a checkbox. You need the raw conversations in natural language, interlinked, with citations and sources, so that a chart answering “what is my team doing this quarter?” can be inspected, assumptions and all. Perplexity listing the pages behind an answer is the same logic. Could Salesforce just build this? Anybody can do anything; Day is a handful of people with laptops. Data migrations are painful, and Netflix and Square have done them, but those moves were from one technology to a relatively similar one. He does not think there is a precedent for the core use of a company’s system of record changing this wildly, this fast.

For an account executive, O’Donnell says, the right analogy is life with a meeting recorder versus without one. The bots are five-star reviewed, and if you read the verbatims it is not about a feature. It is “I can be present in a meeting now.” That is a preview of the benefit across the whole business. People are still clinging to the idea that data entry is cumbersome and should be made better. He calls that 16-bit Pac-Man. Data entry as a concept will entirely go away. What remains is shaping: this part is not quite right, I agree with this take, and the system learns. Administrative work vanishes. You do not write the email from scratch; you look at it and weigh in, maybe as a team, on how you want to sound.

Everything moves up a level, more strategic, more intentional, and you stop forgetting. The hundred-user interview is the same sentence again: things falling through the cracks. Why care? It makes you look bad. Engineering, sales, and customer success all want to belong, to have earned the paycheck, to be respected, not to get fired. The new stack lets people operate at that level of belonging and weigh in creatively on how conversations should evolve, without taking notes. “I don’t have to take notes in a meeting. I can make eye contact. Oh my God, that’s incredible.” Then take it to the next meeting: the details are remembered, the new context is a click, effort goes to zero, presence and self-esteem go up. That, he thinks, is the biggest change.

He has always built sales tools from the buyer’s side of the database. He gets the calls more than he makes them. Buyers love buying things — Datadog, richer traces, Christmas. They also like building rapport with a good rep. The combative posture between buying and selling, he argues, is gunking up the economic engine of earnest people trying to deliver value. A lot of it comes down to trust, and trust comes from keeping promises and remembering details. Sales is already a hard job full of rejection that bleeds into a person’s life and family. Software ought to make that more fun, with no extra effort.

O’Donnell had said out loud as far back as 2017 that he would never build a product from scratch without Pici. When Pici came on the market and ChatGPT captured their imagination, the stars lined up. In retrospect the timing was almost lucky, a little early. In spring 2023 the core trade — take natural language, emit structured output, rinse and repeat — was not yet possible. GPT-4 function calling could be coaxed into JSON. The real unlock was Claude 3.5 Sonnet in June 2024. He still has old prompts from the 3.5 Turbo era, all caps, begging for a type-safe object. Around July 1 he went to New Hampshire for a few days before the Fourth, opened the laptop for 72 hours, and started on pipeline and actions. That is when they let it rip on core CRM.

The wedge was the meeting recorder, the moment of contact, easy to adopt and, it turns out, sticky: roughly 100 percent user retention, now integrated into a deeper CRM. They also spent a miserable first year on Gmail and Google Calendar ingestion, then added Slack. Pici and O’Donnell have long talked about the Spotify of sales: stop choosing a CD and loading eight discs into a visor. Dan Chen, an early user, put it another way. The next generation, watching someone explain Salesforce, will think they are on *Punk’d*. You did the Zooms, the emails, the DMs — and then you *tell the computer*? That button? That form? You really have to fill it out? Spotify had to sign contracts for the music; Day binds to natural-language inputs. Sign up, invite coworkers, auth Gmail, drop the bot in Slack, and within four to eight hours the CRM exists: contacts, companies, deals, tasks.

Deals are the hard part. Company and contact records are a lot of looking at the web, plus a domain and an email address. Opportunity management needs business-process metadata — what a stage means, what “due date” means, the entrance criteria for stage four of the partnership pipeline — and users do not show up reciting that. The system has to pull the metadata out of them, keep re-evaluating as new meetings happen, show proof when it moves a deal, and still let a human make a hard edit. Grady calls it self-driving CRM. O’Donnell agrees, then tells the scare. Automated opportunities landed in November. By December, full self-driving was too scary. From New Year’s on, the push has been the last 10 percent of data quality — fewer false-negative opportunities, better “actions” — and, more than that, a control and configurability layer. See why it is what it is. Correct it. Debug it. Sources and reasoning, the way Perplexity shows its pages.

Internally, Daphne Funston wrote a document called “The Rules of the Game.” Anything the product does, you have to see why the AI did it. You have to be able to override it as a person, and know you overrode it, not watch it flip back. After a user weighs in, should the model freeze the field forever, or take that vote into every future judgment? Usually the latter. Transparency plus control, balanced with automaticity, is in his view the core product problem for every AI-native app. In the lightning round he adds a related debt of thanks to Dario Amodei. For a long time Claude was laughed at for being unusably bubble-wrapped — “I’m done, I’m out” in the middle of a choreographed fight scene. Day now processes huge volumes of customer conversation automatically. That ethical floor is why he can sleep.

Day is a small, high-caliber, remote team. Grady’s firm has been five days a week in the office since 2021; Day is not, and still has flow. O’Donnell’s spoiler is that they will probably end up in an office anyway, fewer sites, fewer time zones, maybe Boston, and the flow they bring in with them will be different. There are no recurring meetings. The remote trick is Slack huddles, low-stakes, sometimes six hours of pair work. Data-model dives are buddy-system, like not going up a mountain alone. The unwritten contract is that everyone holds a huge amount of customer conversation in their head. An engineer who never wants to ping a customer to confirm a bug is fixed will not have much fun. Daphne, a few days in, told him everybody was really nice. Customers talk about hour-scale bug fixes and same-day features, which is how you repay people who are investing time in early software.

When a customer question lands, they write a doc, then pull the code up — when the button shows, what the conditional is. The gulf between customer and code does not need senatorial trappings of go-to-market versus R&D. The last year and a half has been a running update of ICP: personal CRM, tiny-business founders, even starting with VCs as willing early adopters. They stayed with the small set of people who have strategic weight and treat them as design partners who are, in a macro sense, correct. Grady reaches for the Navy SEAL line: slow is smooth, and smooth is fast. O’Donnell admits it is not his usual style. He likes a YC wedge, incremental improvement, broad exposure. System-of-record software is the exception. The investment to get any feedback is high, the scope is wide — they are doing the entire CRM, every function — and there is a completeness bar before you are even in the game, the way payroll is not a one-state launch.

He runs a thought experiment. Race to $1 million ARR and you think in wedges. Race to $100 million and you think about something people do not cancel, a certain ASP, a certain adoption pattern. Race to $10 billion in seven years — the records he looked up are ByteDance in eight, Meta around nine — and every line of code has to make every other line more valuable. Little non-strategic patches go away. The number-one request on the meeting bot is output templates; they have resisted, because templates are a core way of talking to an LLM, and when they ship they will be permanent and correct. You cannot draft an email in this editor without already thinking about marketing email, the knowledge base, the website, the internal wiki.

Two other shifts he is watching. He has been deep in Cursor for more than a year; the coding-tool horse race rearranges every week, and right now less context and less instruction layer often beats a product stuffed with system prompts, especially on 3.7 Sonnet. Hallucination, in their Gmail-classification use case, is much smaller than a year ago; what is missing is tooling for non-deterministic software, which is why he likes the eval vendors. On UX, “consumer grade” in B2B used to mean a coat of paint and information architecture. Anthropic, ChatGPT, and Perplexity hire from social apps, and the bar is now unbelievably fast and unbelievably smooth. Should designers learn to code? He thinks the bigger change is writing: microcopy, flows over time, interaction. Date pickers are a component library. He is not sure anybody will have to code.

In the lightning round, the investor O’Donnell admires most in AI is Sarah Guo. He also wants Dario Amodei on the record for holding the safety line when it was unfashionable. Asked the most extraordinary product outside AI, he names the AGA stove: a hunk of cast iron and a pilot light, no moving parts, four ovens always on, a place on top for wet snow boots, a hearth the whole house gravitates toward. His kids were cooking full meals for grandparents at ten and eleven. The benefit is not a feature list. It is how the house feels.

His Mount Rushmore of product people is also his Mount Rushmore of founders. Steve Jobs is up there, complicated, not a leadership style he wants to copy, with a trust in the user that still sits a little over O’Donnell’s line — he still has to ask people whether they like something. Paul English of Kayak is next, a Boston shadow: the loud, annoying phone in the middle of the room, recruiting that puts an offer in a candidate’s hand within hours, not weeks. Rick Rubin is a Jobs with a different vibe and more humility, the producer of half the records they grew up on, who does not know the knobs or a piano chord and trusts his taste anyway. The fourth face is Sara Blakely, who believed women’s undergarments should not be designed by men who knew nothing about them, did it, and stayed approachable. He had just been telling his ten-year-old daughter that story after she started riffing on maps inside Claude.