How a folder becomes an AI harness with skills and agents

The short answer. Start a hard problem by making a folder and putting everything you know about it inside before you ask a model for an answer. Work you'll want again gets saved as a skill, skills that share one job become an agent, and the folder that holds them is the harness. Once a team shares it, the harness needs two things a folder can't give it: an agent control plane that deploys agents and skills to each person with that person's permissions, and one place where every customer conversation lives for agents to read.

What the video covers

• 5:16 Every hard problem starts the same way: a private-school application gets the schools' websites, every report card and a recorded brain dump, and a house build gets the town ordinances and contractor reviews, all before the first question.
• 9:57 A harness is the folder plus its instructions. Repeatable work becomes a skill when you tell the model to save it, and a set of skills under one job description becomes an agent.
• 11:30 In loop engineering, an agent keeps working until a goal it can verify is met, such as reproducing a bug, fixing it and testing the fix.
• 14:37 An answer can only be as good as the data the model can reach, and for a team that means every source in reach with each person seeing only their share.
• 25:58 A RevOps team already has a repo with skills and scheduled jobs to start from, and a founder without a RevOps team starts from the Claude Code folder they already use.
• 28:33 What to ask of agents deployed to a team: do the agents and skills get better from feedback as people use them?

Excerpts

[0:00] I'm here to talk about the future of CRM, and life that comes after CRM. And I'm making this because I am being asked to, not by my wonderful co-workers or sales prospects, but really everybody in my life. And so I don't come here with all the answers, and I don't come here because I think I'm right about everything, but I'm being asked to talk about it. And I'm making the decision not to write a big script or a think piece, and to just talk. I think about this all day, every day. This and my family and guitars and woodworking. That's it.

[0:30] I don't follow the stock market. I know it well, but it's not interesting to me. It's a distraction. You can tell I don't spend a lot of time choosing my clothes. I'm happy to tell you what I think happens here. Clearly, we're at a turning point with all of these AI frontier models. Something is different now, even than a few months ago, and that was a very different time from a couple of years ago. Certainly, the rate of change is increasing. There's a lot of uncertainty about what's going to happen. I can comment on pieces of maybe all of it, but we'll focus on what brought you here, probably, which is how do I think about CRM as a business owner, as a founder, as a side-project hustler, as the CEO of a Fortune 100 company. I mean, the answers are all the same.

[1:01] It's hard for me to talk about it without also talking about what is going to happen in general, and to try to spot the common themes of what's going to be true in this brave new world that we're entering into. There are a lot of question marks. Am I qualified to talk about this? I don't really know. People are asking me. I'm guessing they're not asking everybody. Why are they asking me? People may widely even think that I'm qualified to talk about this because I've worked for decades. My entire professional career has been focused on the topic of customer data: how to take interactions and details, clues, data, and drive a business forward, and at the right company, to make that customer experience better, which I think deeply in my heart, you will not get me off this. The best way to grow a company is to work your ass off, be patient, stick to your guns, be open-minded, principally root everything in the actual voice, the actual experience of the customer, which now is possible, literally.

[2:04] I think there may be a better reason for me to talk about this, which is my own personal experience with these AI models. The nature of my professional beats and ups and downs is such that I got to start completely fresh at maybe exactly the right time to be, on a personal level, completely native, starting with ChatGPT 3.5 Turbo back in spring of 2023. And since then, I've lived my life top to bottom using these AI models in every possible way that I could try. And I wouldn't say that I am the best in the world at using Claude and OpenAI and the open-source models that you see in the news all the time now. I may believe if somebody told me that I'm the best in the world at using these models in different ways for different purposes. That's kind of my thing.

[2:38] I've obviously gone very, very deep on the software engineering sort of tip of the spear, and I'm happy to talk about what we're doing. […] It's wild. But I use this stuff for everything in my life. Without really overthinking why or how it fits into my life purpose, it just was kind of happenstance. I started an AI-native startup in May of 2023 and started coding, started building, started raising money, doing finance stuff, everything from scratch, just with a couple of friends.

[3:41] And then I translated what I was trying and what I was learning into my personal life, my family. I translated it then pretty quickly into my hobbies, my creative work. I take guitar playing pretty seriously, which is not to say I'm any good, but I like it. I like woodworking. I like building things with my hands. I like cars. I like power tools. I like art. And I have a lot of friends that are creatives: songwriters, musicians, artists. My mother's an oil painter. Architects, you name it. These are the people that I'm really drawn to, and these people have really struggled with how to think about it. Educators. And I've spent a lot of time in dialogue and fellowship exploring those topics.

[4:12] And if we want to talk about customer data, I find it easiest to enter this topic by completely forgetting about work, forgetting about CRM, forgetting about customer data, and to try to spot the big change that has happened in the world. The advent of these models, the idea that they could create things for us, with us, the advancement of those capabilities, the judgment, the precision, orchestration coming into view, things happening asynchronously, things happening automatically, waking up to piles of work product to review. Originally, that was not very good in any domain that you were working. The idea of agents, skills, harnesses, context graphs, all of these amazing application-layer, so to speak, innovations that have become conceptually appealing, but for many people not yet concrete. And that gap, I think, is what I can help close, by spotting what really has changed in a simple way.

[5:16] Here's how I think about it and how I would describe it today. If we have a hard problem to solve, just as a human, if you have a hard problem to solve today, a big interdisciplinary, creative, technical, financial, whatever problem, we are all approaching it now, we're starting to, in the same general way. And the nature of that, I think, is the right entry point. We start now by deciding it is a big hard problem to work on, and creating a home base for it on our laptops, creating a folder.

[5:47] The hard problem could be planning a wedding, dealing with a family medical issue, building a house, learning to play an instrument, getting a new job, getting promoted at your job. Anything at this level of sophistication, the right starting point certainly is to dedicate a space to it and create a directory, a folder on your hard drive. And then what do you do? You put into that folder all of the information that you have on the topic. Let's say you're trying to get a kid into a private school. What do you do? You take all the target schools and you convert their websites into markdown, and you clean that up. Why? Because you care about these schools. You care about what they're good at, who they're for, or you should, everything about them, where they are, every policy. You don't know what matters yet. You get all of the information you can, and you put it into that folder.

[6:18] And then you take every report card as far back as you can possibly get them, every teacher's note, any detail that you have laying around, you put those in there. And then you turn on your voice recorder, transcriber, Wispr Flow, whatever, and you just riff. And you take all of that context that you have personally, and you get it out, and you turn that into text, and you put that in there with an explanation of what it is at the beginning. And you do all of this before you try to solve the problem.

[6:50] We no longer are taking a school application and the questions on it and pasting it into ChatGPT or Claude and saying, help me answer these questions, or help me fill this thing out. You start by really trying to master the problem, or at least provide all of the relevant details, so that a frontier model can further clean them up and make connections and map out the problem space for you. Spoiler: you're going to write the answers on that application. That's your job. You are not going to copy and paste this stuff. But first, you need to really understand what the truth is.

[7:20] You got to figure out where this kid is supposed to go to school. You're supposed to figure out which wedding venue is the right one, based on the weather and partner's preferences, and your wonderful future in-laws, and it's possible they have some opinions. I mean, you got to document everything. Building a house, think about it: budget, timeline, zoning, town ordinances, the easements and this and that, interviewing general contractors. These are the things that we do day in and day out in our lives. These are hard problems. We're adults. We have hard problems. They take more than 5 minutes. They take more than an hour.

[7:51] You put everything into one place, and then you begin to understand the nature of the problem. Then you begin to draft, and poke holes, to test, to simulate, to imagine, to sleep on the best solution you have in mind. Write down all the options, including the bad ones. Play devil's advocate. Play them all out. And from there, we start to get answers that also live in files on our laptop. And those answers are pretty compelling, and those insights are pretty useful. And then what do we do? We have to do something with it. And that will involve usually some level of human interaction at some point.

[8:22] We still have to go to the bank and apply for a loan. We still have to call our future mother-in-law. We still have to tour the school we're applying to. We still have to do the hospital visit with a sick parent. The question is, how prepared are we? What are our standards for how we comport ourselves and how we spend our time effectively? Are we able to ask good questions that fill in the missing details? Are we able to tell stories in ways that land, that make sense to people?

[8:54] And now we can start to take this into the workplace. For every hard problem, or hard set of problems, type of problems, maybe for every role, it's no different. We create a folder, and we put the information in there, and we start to draw connections. We start to draft solutions. We take it as far as we possibly can. And then we learn a very interesting thing, which is that we can actually add capabilities to these AI models. If they need to see things, we give them ways to see. If they need to hear things, we give them ways to hear.

[9:25] Make a meeting bot that can go to a meeting, take a transcript with consent, and listen to every single word. Write it down, and maybe figure out who said what, and then maybe who works where. But why does it matter that this person said this? What does it change that that person said that? As we add these capabilities to these projects, these folders, these harnesses, effectively that's what you are building each time you tackle a problem or a set of problems by creating a folder. That's what a harness is. That's it. It's not a super complex thing.

[9:57] It's a set of instructions, things that you do that work that you want to do again in exactly the same way. You save that as a skill. If you find that set of skills really fits under a particular kind of general job description and way to think about approaching the performance of that work, well, that becomes an agent. And these are files, and they live in a .claude or a .codex or whatever little hidden folder inside that project that you've created.

[10:29] And these become quite powerful very quickly, because we learn that those are not things that we have to write. They're things that we have to notice are worth saving. Again, these are the human ideas that creep in. And so we tell Claude, "Hey, that was really great, the way you edited that video, the way that you found that thing online so quickly. We're going to need to do that in the future. Save that as a skill for yourself, based on everything you learned." Okay, done. And then we have a whole bunch of them.

[10:59] We say, "You know what? I keep giving you feedback that is across all of these different tasks. I want you to collate all of that and create a video editor agent that knows how to do all of these things that you've learned how to do, that you've documented how to do in the right way. I want you to create an agent, a job description, the job training in it." And then the next time we go to do that type of work, all that stuff gets pulled up automatically, and the work is done well. Defined by our standards, defined by the human standards. And then we learn that we can do even much more powerful things.

[11:30] Maybe we've added enough capabilities so that, for a particular task or goal, the AI is now left to work autonomously until the success criteria of that goal are met. To do that, it has to be verifiable. If you want the AI to fix a bug for you, you have to give it a way to reproduce the bug, to write code, obviously, and to test it to see if the bug is actually fixed. This is huge. Absolutely massive. Once we adopt this way of thinking, we're doing what's now called loop engineering. And that's a big leap. It's bigger than the leap from loop to graph engineering. Controversial, but so what? Graph engineering is just map-reduced loop engineering, if you ask me.

[12:02] These are big leaps that happen. They are not as big as the leap of creating a place to hold all of the state, all of the instructions, all of the progress and the breakthroughs, for future work to be higher quality, faster, less expensive, more consistent, more useful. So you can see very clearly that we are not copying and pasting into web UIs anymore. We are building systems that will wake up in the middle of the night, that will respond to events, that will check on things for us, that will watch things for us, that will listen, not just for a search term, but for patterns.

[12:34] As a CEO, anything in the world that you want to know or want to watch, including when you don't know what to look for, you can say that and have that be what you wake up to every day, or the reason that an agent reaches out to you to give you a heads-up. The things that I've been able to do with these very basic concepts, not technologies: I could have a brand-new laptop, nothing on it, and I can use this stuff in wildly powerful ways that would blow your mind. I don't care who you are.

[13:04] What is the role of CRM in this new world? I'm sure you've already figured out what I'm teeing you up to see, which is, well, wait a second. If this is a team problem, what do I do? It's the right question. Okay. So we can use the CRM as like a database and have that maintain the state. Sure. Absolutely. And how far does that get you? Well, for any of the problems that we've talked about, it's immediately obvious that garbage in, garbage out. If I'm building a house and I put into that folder a bunch of crayon sketches and nothing else, these models will do absolutely everything they can, but I'm not going to get a very good result.

[13:34] If I put in all of the town building codes, and ways of accessing all of the suppliers dynamically over the API that are in my area, and ways of finding general contractors and subcontractors, and finding reviews for them and instances of work they've done; if I invest in the context and the capabilities within that folder, within that harness, that project, when I say that I need to do it for $100,000 less, and I need better energy efficiency, in one of those scenarios I'm going to get a really good answer. In another scenario, I'm not going to get a very good answer at all. Or worse, I'm going to get an answer that looks pretty credible and in fact misses the entire point.

[14:05] I can sit there all day and do a better and better job of saying what I want, and the resolution of the answer is going to be limited by the resolution of the data that it has to work with, that has been ingested, that is sitting there available. The more I create capabilities for a system to build, to research, to experiment, to verify, to simulate, to evaluate itself, that's when the results start to get spectacularly good.

[14:37] In my day-to-day life, it is essentially never the case that somebody has to call a one-on-one with me to bring me up to speed on something customer-related whatsoever. I have an agent that exists to find the gaps of what I know about and to fill those in, to research very deeply, to consider what I already know, to look at the priorities, the challenges of the business, the student body of my colleagues that are serving customers, every detail about every one of these companies we're trying to serve, and to say, "Hey, Christopher, this next part matters."

[15:38] For this to work at a company level, we need access to everything, and only the parts of everything that each of us is allowed to see. This is a very hard problem. It's a hard problem even when you're just dealing with data that people have written down by hand, typed into some web app. It's a monumentally difficult and unspeakably high-stakes problem when the data you're working with is primary source material. This person said this thing at this time. This person wrote this thing in an email at this time. You need to be very careful, and it needs to be extremely fast and extremely efficient.

[16:10] People talk a lot about the cost, and the token cost, and all. Yes, that's absolutely huge. The real cost of those inefficiencies, though, is time, especially if you're still waiting for this stuff to happen because you're asking a question ad hoc. It needs to be unbelievably fast. And the quality of the answer: if a model has to sift through a bunch of thin, manually entered data to try to find something that matters, and to try to reverse-engineer the story and the context behind this. Why is this deal in this stage? Why is this person marked as the decision-maker, the economic buyer on this deal? Or do you even have that data?

[16:40] Now it starts to become clearer what the limitations of a legacy CRM system are. In many cases, the data that is most relevant to the task at hand, it could be a question, it could be an email draft, you're trying to do whatever it is, a report, that data does not have anywhere to live in the data model. It's not data that anybody would be entering by hand, unless you had 142-IQ temps that never slept. It's not realistic to get the data without AI. And in the legacy CRM systems, there's just nowhere to put it. And so it's not even about how fast it is to retrieve it. I mean, it's not a thing.

[17:11] And one of the reasons I'm making this is, transparently, just that business leaders, go-to-market people, sales, marketing, customer success, you name it, are really kind of too late to the party. This is what it looks like when I put my foot down, I guess, and I'm going to shake everybody by the lapels and say, "Look, the software engineers have flying cars. Lawyers have flying cars now. IT has flying cars. Finance is starting to have flying cars. And sales and CS are just manually inputting stuff, and totally tolerant of these ridiculously avoidable gaps."

[17:42] What's our plan for this renewal? Holy, we have like a week. This is an open-book test, man. All of the data, all of the information, powerful models to comb over it and come up with a plan. Everything that you're deciding in management meetings, ops meetings, everything. This is all capturable, and can be put together in one harness in a way that makes it unbelievably easy to solve these problems, or at least to be making progress that builds on previous progress. It's very easy to do that now. And I do believe that go-to-market teams are very close to having this epiphany. And when they do, I want there to be a really credible answer to what happens to CRM.

[18:41] I gave a talk at Inbound […] and all these fancy new product releases, and that was great. I don't really dwell on any of that. I keep coming back to crossing the street in the Seaport, and this wonderful middle-aged woman coming up and stopping me in the middle of the street. There wasn't any traffic. And we had this conversation in the middle of the street. She stopped me, and she said, "I just have to tell you the story of the last year of my life. I was in this toxic job. I was commuting 45 minutes to an hour each way. I'm a single mom dealing with pickups. I'm dealing with trying to help my kid as she becomes a teenager. I mean, I was underpaid, and more than that, underappreciated." Spoiler, that's really what matters.

[19:13] And so she said, "I learned this product that you and your friends have built. I learned it, and I put that on my cover letter. I said, I am capable of doing the following things. Put me in, coach." And she said that she found a job 15 minutes from her house, small light-manufacturing company, like outside Rochester. She's totally present for her daughter. She makes a bit more money, good bit more money, and most importantly, she likes the management team, and she walks on water to them.

[19:44] That's the kind of impact that we can have with this stuff. I mean, you can run a business, but even more importantly, you can feel comfortable at work. You can feel like you're doing something of service, of value. And ironically, with the AI, you can absolutely feel as though you're doing things in a way that nobody else could. Why? Because you're spending your time on creative things, on matters of taste and judgment. For every hour spent not having to be brought up to speed by sales reps at my own company, we spend that hour thinking on a human level: should I text this chief customer officer at this prospect that's on the one-yard line, or should I wait? What's gracious? What's a right-size kind of approach to this deal?

[20:17] That's what I can sit there with Sam, our sales rep, and we can have these very thoughtful conversations about how to play things out. He comes into the office, he puts his knapsack down. His expectation is, I'm up to speed. I know who this person is, who that person is. This extra security review that's coming in at the 11th hour, we know all about it. Engineering knows all. Everybody knows all about it. Great. Yeah. Okay. So we're going to get through that, and then this, and then that. And then, marketing side, website updates, customer copy, all this kind of stuff, quotes, testimonials. We're all working off of the same data set.

[20:48] And it's trippy, and it's, importantly, a lot of fun. And it means that we can sit in fellowship without staring at a computer screen, and we can focus on getting honest with each other and coming up with plans, making hard decisions. I got a document from a sales rep the other day that actually led to this video, and came with a bunch of quotes about questions people have about life after CRM, the post-CRM world. And so I'm not in here because I think it's a great idea. I'm in here because you think it's a great idea. That's pretty cool, man.

[21:20] I mean, it's a lot better than sitting in 9-hour QBRs looking at charts. We all know the data. We get the data. The data is scrubbed. […] I don't ever show the board something that hasn't been triple-verified. If I say that we did this many demos, those are actual recorded things where somebody shared a screen, as though somebody had gone through Gong every meeting, watched every single one of these recordings, verified everything. What does it cost me? Almost nothing. So I can spend my time in a board note talking about what it's like selling and servicing this, where the world's head is, asking for help.

[21:51] This is a wildly different context, because the way we solve any problem now is just a completely different pattern. What was the old pattern? What were people doing in, let's say, sales, before the models, and before we had this kind of harness and this kind of data back end, this kind of substrate that gives us these superpowers? It was enough. And even today, for a lot of companies, that's still the expectation. It is enough to take a few notes, have a cheat sheet of who you've been talking to, and maybe, if you do what your manager tells you and you're diligent about keeping the next steps field on the opportunity up to date. […]

[22:51] It used to be enough that you could collect business cards, keep them in a stack. We get the internet: let's put the spreadsheet up on the internet. The assumption was that the human is effectively going to solve all of the problems, is going to be the source of truth for any insight into the business at any level of granularity. That's going to come through a human, if it ever makes its way onto disk. And just having somebody who is reasonably good at capturing this information and reasonably good at doing something about it. Geez, I think I'll call Julie again today. See if I get her to do the thing, whatever. Moving these deals along.

[23:22] That has been enough to justify paying somebody six figures, in the enterprise seven figures on a good year. And that all absolutely goes away. Absolutely goes away. It means wonderful things for sales reps specifically, because we still need representation. We need a human to be the UI. We trust names and faces. We connect with personal history and tastes. As a buyer, I love sales reps. I've always loved sales reps.

[23:53] My dad was an equity analyst, and he raised me saying, anything you ever want to know costs a nickel. And what he meant by that was you could call the sales line at any company, and they knew everything about their product, about the industry, about the company, leadership. […] I have been shown slides, and not every slide, only when they're really good and challenge my worldview and have a real insight in terms of how to talk about all this stuff that we're doing in our own business.

[24:24] And every time I see one of these slides, it's beautiful. It's completely on brand. I mean, these are not hard problems. I started seeing these the first, second day somebody joins the company. They're able to do this. I don't even have to train them. They just pick this stuff up from their friends, and you're good to go. Everything that people come to us looking to do is now so trivially easy to do. It blows my mind. Activity reporting: you want to know who's making calls and who's sending emails and who's doing the cold outreach better than the other person. Template reporting, whatever. So easy to do.

[24:55] You want to forecast. Okay. What's your opinion? What's your sense? What do you want to explore within this? Are you interested in how to close the gap between the base case and the best case on the forecast? Do you want to know, as CEO, exactly what you can be doing? Now we're starting to get into the territory of where it's interesting. But there are entirely new lenses now on the questions we ask, the things we watch. If you want to see how senior somebody is at one of these information companies, these laptop jobs, cubicle jobs, whatever, if you have the right harness, you can go in and you can just ask to see the decisions that they are making. You can ask to see the promises that they're making, and whether those promises are getting kept. And you will within seconds be able to see the quality of the work on a human level: the work that remains, the decisions, the promises, the trust, the intuition, the compliance, the innovation.

[25:56] The way to think about CRM, and what the stack should be for your company, is to start with whatever folder has been made. At a more established company, RevOps has one, guaranteed. They call it a repo, because they share it through GitHub. It has a lot in it. It probably has an entire web app, or multiple web apps, for interacting with the data. It has scheduled jobs that run in the cloud to grab Slack messages or Gong transcripts, maybe emails. It has ways of accessing the internet. It probably has a skills directory, and that team is probably meaningfully along the way toward getting all of that to run automatically, thinking about how to automate some of the work that people on the front lines at that company are doing every day.

[26:29] For people who don't have a RevOps team, you are RevOps. For founders looking for their first CRM, I urge you to forget everything you know or think you know about CRM, and to start at the very basics. Maybe you have one of these already. You have a folder that you're using with Claude Code, and it has all your notes and decisions and your product roadmap, persona definitions for product-market fit, all that kind of stuff. You're trying to work up the sales process. You want to try to run it yourself, then hire some go-to-market people, scale this thing, do a Series A, get to Series B, etc.

[27:00] Start there and play that tape forward. Where does that all go? Where are you getting the data that you're basing these decisions on? When you do a customer interview, does it update everything? Does it change your mind? Does it reach out to you? If you and your co-founder are off doing all kinds of different things, and then your website traffic behavior changes, and this and that, do you know? How quickly do you pivot? How quickly do you update the copy? Because the way you're talking about what you do is different from the way that the people you're on the phone with talk about it.

[27:31] If you keep that daydream going and you imagine the absolute perfect company brain that you would build, you're going to have two big areas that are appropriate for a vendor to fill. One is agents, the agent control plane. When it's just you, you can mostly get by with Claude Code or Codex agent definitions, skill definitions. You can mostly probably at that point run them automatically one way or another. Certainly on your laptop, if nothing else. You could have something that will run in the morning if your laptop's open.

[28:03] As you start to scale, if you want a new hire's agent to be able to see parts of your email, where you want to control what meeting transcripts are shared with a growing team and which aren't, these are vendory type of problems. Full stop. As the team grows more and more, giving people AI agents is something that can be done properly, and something that can be done really in such a hodgepodge way that there's no benefit, there's no adoption. One of the things to look for there is: are they adaptable? Are they coachable? If you give your company a whole bunch of agents with skills, do those agents and skills get better over time just as they're used? If not, you need a different architecture.

[28:33] You can take the agents and skills and all the work that you've done. That's fine. You're going to need a way to have self-improving agents. Not even self-improving, but agents that just do what they're told. The thing on the go-to-market side, it's different from engineering, because in engineering, you can have a front-end engineer agent, or a database performance agent, and pretty much there's one way that the company wants to solve a given problem, and that can be recorded in that agent and its skills. In go-to-market, it's a pretty different story. You might have one sales process. You probably should have one set of guidelines for how to update the CRM. Okay. Well, now you need those to be followed perfectly. And you don't want to spend anybody's time on that. That should be almost free at this point. […] Eliminate all that time out of a seller's calendar. Should be gone.

[29:36] And that data entry should be something that can be evaluated and improved episodically or dynamically. And you start to get a compound growth curve to the quality of your own harness, which is operationally the beating heart of your business. You will see in this new world, and the list goes on and on. But deploying agents in some sort of agent control plane becomes necessary. Arguably, at a two-person company it is already necessary, but certainly as you start to scale and take financing, etc. Absolutely necessary to be an adult and to get the kind of results you need.

[30:07] The other big area that you'll run into is: where's this data live, in what form? How much work and integration complexity, operational complexity, tech debt, etc., IT glue? How much of all that exists just to get data into a place that your agent can read it? […] For the core of what you're working with, it needs to be in one place. And that place needs to be lightning fast.

[…]

[32:15] The new data substrate that will become what we all live and work on, off of, on top of, surrounded by, every aspect of our work. Instant access to all of the relevant things that have happened, words that have been said, decisions that have been made, processes that need to be followed, everything. Instant, complete awareness of everything that matters. This is a completely new data substrate. It's called a context graph. It is not something that is to be built casually.

[32:46] It is not something that will be built by doing the intuitive, oh, we'll just do it this way. No, this takes real skinned knees, trial and error, and some really talented people working on it. Few of them, preferably, and they need to at some point along the way be right. And then they need to pull off the build in a way that actually works in production, and not just on someone's laptop. That's a lot of gates to clear. […] The need for this data substrate has been obvious to the world for at least 9 months. December of 2025, Foundation Capital wrote a piece called "the trillion dollar opportunity" […]

[33:18] It's hard to build. It's easy to use. If we go back to the way that we solve any problem today, in this day and age, we're going to need all of the information instantly, in ways that are secure, fast and efficient. Meaning these agents, these AI models, can get exactly the information that's relevant, excluding the stuff that's not relevant, […] in a form factor, in a shape that allows it to dig deeper, to make connections, and to run around the entire history of everything that's happened, and find all the things that matter, and then answer or take action.

[33:50] And once you touch this and see this, you will want it running around the clock as much as possible. And the way to make that happen is to define virtual workers. Even if they're in the background, they don't need to be on the sales calls. You should be on the sales calls. You should be talking to the customers. You get a very different result. This isn't about deflecting support tickets, figuring out which knowledge-base document to send to somebody so that you can close the ticket and claim victory somehow. This is about detecting that something doesn't work the right way for a customer because of two Slack messages and this comment on a meeting recording and this insight from somebody in support, in a team meeting, and then just fixing it automatically.

[34:48] I teased a little bit about how we do the software engineering at Day AI, but that's how it works. Anybody at the company can say anything in Slack and have it generate a pull request, a unit of work with an answer for how something works, or with an investigation of a performance optimization, or new design options for something. And if we like it, it goes to production. Simple bug fixes go from customer report to production. And we're starting to identify these issues. In many cases, we're able to identify an issue before the customer even knows it happened, or certainly way before they're able to submit a support ticket. The fix is already building in production by the time they go to write us. That's incredibly cool.

[35:22] And I think you can see why, when we talk about the future of CRM, we're actually talking about it through the lens of life after CRM. […] And we now live in a time where extremely difficult problems, like running a company, can be solved, ongoing, optimized, the benefits compounding, progress being made token by token, bugs fixed, sales processes improved, reps trained, prospects found, prospects writing back to emails because the email was perfect and spoke to them in hauntingly specific ways, because that prospect looks like a handful of other prospects that had a common shared positive experience. And so that cold email is no longer bumping this to the top of your inbox. The cold email is: I have a real solution to a problem I bet you have, and I have expertise and credibility. I have relevant reference customers. I have everything I need to get you on the phone.

[36:23] There's a big topic in the news. It's faded a bit, but it is a very real one, which is: are people just going to build their own CRM? Sure, absolutely. Permissioning classic CRM data is still kind of a gnarly problem, but it's solvable. Somebody will do something open source, I'm sure. And you can have something that you run yourself on Vercel or Netlify, uses Postgres in the cloud, and is as good as any of those things, and your data is completely, totally private to you. That's fine. That's not that hard to do. That's not the problem. […] There is a class of enterprise value that is now available to unlock that is so far outside the expectations of these go-to-market teams. It's even outside of what they believe is real or possible, and there is no way that that will stay unknown and hidden forever. There's no way. It's too good for the people using it, and it's too good for their prospects and customers. […]

[37:26] Day AI integrates with these legacy systems, and when people go to do these integrations, often they don't bother. That's the truth. The truth is that it is very, very useful to delay moving from a system of record to some newfangled context graph thing. I get it. I get it. Look, I'm probably one of the few people that has moved a public company from one CRM system to another. So, man, I know what's involved. Boy, I kind of wish I didn't. It's super gnarly. And you don't have to bite that off on day one.

[37:56] But once people get in and they see the power of context graph, they see the power of living agents, agents that are truly in a governance plane and truly deployed in the right form factor, and you see the depth of the data that can live in a context graph, you realize that there is nowhere to put it in the legacy system. It is a superset of the legacy system. You cannot ever play Elden Ring on a Game Boy. And if you're thinking about playing Elden Ring on a Game Boy, even if it were possible to port it in some way, the problem is that you're still thinking about the convenience of a Game Boy. It's really sobering to see how big the change is.

[38:28] The good news is it is absolutely incremental. You can grab a system like this. […] And I think that a year from now there will be a lot of things that look very similar to what Day AI is today. And it's up to us to have the next thing. And that's fine. That's what gets me out of bed in the morning: building stuff with my friends, bringing it to customers. That's the best.

[38:58] It is shocking to see how different it is. And it is challenging and scary at times to think about how much it changes a business. The good news is it's also invigorating. It's very good for the careers of the people who decide to take it on. And since this is a concept that extends and bulletproofs, accelerates, secures the company brain that you probably already have, there's very little risk.

[39:30] So what is life after CRM? Life after CRM is roles being combined and people having real agency at work. It's a prospect who reads a blog post with a name and a face on it, then downloads a guide, some valuable thing, a context file that was made by that same person, then gets an email from that same person, that decides to book time with that same person, buys from that person, talks to that same person the entire time they're using it, does a renewal with that same person. Probably doesn't have to write support very often, but when they do, they are hearing from that same person. […]

[40:01] They are able to sit in fellowship on a Zoom with a customer they know very well, to have strategic conversations, to develop their own business acumen, to be sharing patterns that are working, or maybe hearing about a new pattern and then very quickly turning that into marketing materials, getting more leads. Rinse and repeat. This is very good for the buyer and very, very good for a vendor […]

[40:31] Certainly, life after CRM becomes […] a bigger W-2, longer, deeper customer relationships, greater likelihood of the startup you chose to join succeeding, more choice in where you do go work, especially if you adopt this way of solving problems earlier rather than later. It's my pro career tip. Life after CRM is pretty chill. Running the most ruthlessly optimized self-healing business ever seen on the planet will feel pretty interesting and chill. More and more marketing leads are going to come through word of mouth for companies. It's healthy. Life after CRM is pretty great. And you could have that life.