People keep telling me I should say in public what I say in private, so … here we are.
Somewhere in your company, right now, someone is closing out a deal. Nine months of calls. A champion who stuck her neck out for you internally. An objection in March that nearly killed the whole thing. A Slack thread nobody outside the deal team will ever read. And at the very end of it, someone opens a dropdown and clicks Closed Won.
That’s all the company remembers.
Everything else, the hesitation you could hear in someone’s voice, the exact sentence that finally got the budget approved, the reason it almost died in March, gets thrown away on the way in. Thrown away on purpose, by the software, because there was never anywhere to put it.
That is what a CRM is. A database built to take the richest, most human information a company owns and boil it down to whatever fits in a field.
It is not going to exist much longer. The premise underneath it (that a person sits down after the fact and types in what happened) is already gone. Software watches everything now. Models can do the hard work of running a company, but they cannot run it from key/value-based data.
I have spent twenty years on this problem space. I built what I believed was the best version of the old machine, and for the last three years I have been building what replaces it. I’ll tell you how I got from one to the other, and I’ll tell you up front that the answer indicts an industry I gave my best decade to. It doesn’t spare me either.
Riding my BMX around the neighborhood as a kid, I did not dream of growing up to spend decades building CRM software. Nobody does. It turns out even the most open-minded high school girlfriend eventually stops wanting to build a family with a struggling guitar player slash low-stakes poker player. The web had shown up when I was thirteen, I started building stuff immediately, and after college I cobbled together poker buy-ins in my mom’s barn by making websites. Somewhere in my mid-twenties (yes, I ought to have done the math by then) I figured out I could build actual products instead of websites …
A run of improbable luck later, I was leading product at HubSpot … from age thirty to forty.
I walked in with one idea I’d been chewing on since my first job, a marketing associate gig where I ran web analytics and babysat an ecommerce shop and wasn’t senior enough to touch the marketing emails. I knew how I wanted to handle multiple form fills off the same cookie. I knew forms, landing pages, contact lists, and automation should be one thing instead of four tools that don’t talk to each other. With this exceedingly random hand, I flopped a straight flush draw: the shot to rebuild what became Marketing Hub, then build Sales Hub from nothing, then stand up teams under leaders of their own for Service Hub and Data Hub.
There was one thing I said to every single person I ever recruited, probably ten thousand times, about why any of it mattered:
The more tightly a company is bound to the ground truth of its customers’ real experience, the faster it grows.
I have always believed this the way we believe a law of physics, there’s nothing to believe it’s just true.
A chronic, incurable disease made me leave HubSpot - luckily it is in remission, at least for today. That is a story for another day, but I’ll say this much: the guy who held that role for ten years is gone forever. I statistically shouldn’t be here. I am though (I think?). And when you get time you weren’t supposed to get, you spend it differently. No more performing for people as a character (“CTOD”) and started being myself, Christopher. Today I am trying to get out of the way of what the Universe is trying to do, that I need to do the work to let happen.
I left at the moment of the all-time stock high, and by the time I got my health back, the role was gone. I lost my life’s work, and it has been musical chairs in that job ever since. It’s heartbreaking. Maybe. Maybe … who gives a shit? What matters is forging the future and bringing the next-generation answer to the world.
Building an all-in-one “CRM Suite” was a beautiful problem. I liked the people. I was learning faster than I ever had. Some part of me knew I’d been handed a Super Bowl shot I hadn’t necessarily earned. I wanted a win, and needed a white picket fence. Under all of that there was a hole I kept trying to fill by being excellent at something, anything, chasing it manically, not even toward a goal. Just going.
For a while that hunger needed a villain, so I told myself Salesforce was some lazy old vampire of a company that couldn’t be bothered to build anything itself. Give me a break. They were good, smart people doing good things for good reasons, same as us. We weren’t better. We were, I’ll grant, better looking, and better cooks, and better musicians. Coincidence, and a meaningless one.
And the thing worked. We were the Mac to Salesforce’s PC. We took the message Salesforce had already put into the market and delivered on it, and the business ripped. I mean RIPPED. Then I was gone, and it stalled.
For a while I had a comfortable explanation. The COVID bubble burst on everybody. Fair enough. But the bubble passed and the stall stayed, and I couldn’t put down the question underneath it. We had done it the right way. One platform, built natively, no “tuck-ins”, no bullshit hand waving … live working software on one platform or get the fuck out. The market bought and held and bought and held for a decade, customer satisfaction steadily rose, it all felt like one team: employees, shareholders, customers.
Then it stopped. Why?
I’ve asked many friends who are better “market analysts” than I am. I’ve read up on the Consolidation Curve, the endgame of a company (or industry’s) maturity cycle, some smart folks have offered reasonable explanations and I can try to distill that for you, to the best of my ability.
CRM didn’t mature the way I always assumed software matures, by getting steadily better at the thing it claims to do. It matured the way oil and cable and railroads mature: by consolidation. Once you own the customer database, the playbook writes itself. Buy the marketing automation company. Buy the analytics company. Buy whatever your customers keep asking you for. Collapse five software budgets into one master contract and go upmarket as fast as your sales org can drag you. Wall Street calls that maturity, a category settling into its natural oligopoly. Anybody who has written the code, or carried a number may say it’s architectural decay with a rusty crown. Gen-Z Dickens fans will say it’s giving Havisham.
The consolidation phase. Building it the right way was only a really good answer for earlier phases of the maturity curve.
Oof. So that leaves customers with a handful of products nobody particularly loves, run by companies that get less customer-centric every year. HubSpot, whose whole identity used to be “Get Shit Done” mostly stopped building and started buying. As a builder, and with a heavy heart, I have to ask … where did all those products go? Meanwhile Salesforce kept being Salesforce, more like Oracle every year, the young Vader who admired the Emperor slowly becoming him. Lightning coming from his fingers (oof dad jokes), Sales Cloud still running literally ON Oracle (fun fact).
Maybe that makes a decent essay about consolidation. It’s not the whole story. Consolidation might explain why nobody fixed the problem but it doesn’t explain what the problem was.
To see that, I had to stop looking at the industry. I had to force myself to forget everything and start over on a blank page. I was pretty sure I was done with software forever.
In 2023 I started over. I didn’t need the money, and I didn’t have any need to be known for anything (btw for anyone who wants to be famous: a) you don’t and b) software is a bad way to pursue that), which is why I’d rather be heads-down with a few friends building software and talking to customers than writing this.
Mostly I just wanted to build again, and I started completely fresh, at what turned out to be exactly the right time, with the first models that could do anything useful.
Since then I’ve lived my whole life inside them. Work, family, hobbies, everything. And somewhere in there the way I solve any hard problem changed. I want to show you the change without thinking (for a minute!) about “CRM”, because that is the best entry point in my opinion.
Say you’re trying to get your kid into a private school. Two years ago you’d have opened the application, pasted the questions into a chat window, and asked for help with the answers. Nobody who is any good at this does that anymore.
You make a folder on your laptop. You take every target school and convert their websites into markdown and clean that up, because you care what these schools are good at and who they are for, and you don’t yet know which details will matter. You pull every report card as far back as you can find them, every teacher’s note, every scrap of paper with your kid’s name on it. Then you turn on a voice recorder and you talk. Everything you know about this kid, what she’s like at the dinner table, what she’s scared of, what lights her up. You transcribe that, put a note at the top saying what it is, and drop it in the folder.
You do all of this before you try to solve anything.
Then you begin, and you begin by understanding the problem. Which school is right. What the gaps are. What is true. You draft, you poke holes, you write down every option including the bad ones and argue against each. You sleep on it. The answers that come back are good, and they live in files in the same folder, and they build on each other.
And then you write the application. That is your job. You are not going to copy and paste. But you write it with mastery of all the relevant details, and what you write comes from the heart because you know you mean it.
Same thing if the task is choosing a wedding venue and the weather (sadly unpredictable) and your future in-laws’ opinions (sadly very predictable). It is the same with a sick parent and the hospital and the insurance. It is the same with building a house, where the folder holds the budget, the timeline, the zoning, the easements, the town ordinances, and the notes from every general contractor you interviewed. These are the hard problems of adult life. They take more than five minutes. They take more than an hour. And the way we approach every one of them has collapsed to the same approach: make a dedicated space, put everything in it, understand before you make any decisions or take any actions.
At the end there is still a human interaction. You still tour the school. You still call your mother-in-law (good luck!). You still sit in the hospital with your dad. The only question is how prepared you are when you get there.
Now take that folder to work.
For any hard problem at work, or any type of problem, or any role, it is no different. You make the folder. You put the information in. You draw connections and draft solutions and take it as far as you can.
Then you learn something interesting: you can add capabilities. If the model needs to see something, you give it a way to see. If it needs to hear something, you give it a way to hear. You build a little bot that joins your meetings and writes down every word and works out who said what and who works where. Then you ask the harder question. Why does it matter that this person said this? What changed because that person said that?
Every time you tackle a problem by making a folder and adding capabilities, you are building a HARNESS. Look at you go. THAT is all a harness is. A set of instructions and a set of things that can do things. When something works and you want it done the same way next time, you save it. That is a skill. When a set of skills clusters under a particular kind of job, with a particular way of approaching the work, that becomes an agent. These are just … text files. They live in a hidden folder inside your project. There is nothing exotic about any of it.
You don’t write these things yourself. You notice them. You tell Claude, hey beloved robot … that was great, the way you edited that video, the way you found that thing online so fast, we’re going to need that again, save it as a skill. Done. A month later you say, I keep giving you the same feedback across all of these tasks, collate it and make me a video editor agent that knows everything you’ve learned. It writes the job description and the training. The next time that kind of work comes up, all of it loads, and the work gets done to your standard, because your standard is what got saved.
Now we leap, bigtime. You’ve added enough capabilities that for a particular goal the AI can work on its own until the success criteria are met. For that, the goal has to be verifiable. If you want a bug fixed, you give it a way to reproduce the bug, a way to write the code, and a way to test whether the bug is gone.
Then you leave.
We are not copying and pasting into chat windows anymore. We are building systems that wake up in the middle of the night, that respond to events, that check on things and watch things for us, that listen for patterns nobody thought to search for. As a CEO, anything in the world you want to know or watch, including when you don’t know what to look for, you say once, and it becomes what you wake up to.
None of this is a technology. These are ideas. Hand me a brand new laptop with nothing on it and I can use them in ways that would blow your mind, and I don’t care who you are. I could blow Boris Cherny’s mind. I could blow Thariq Shihipar’s mind. (I was his first boss, I get to say that once a day max. Hey bro, nice work out there.)
Half of you reading this have already built yourself a version of this. Notes, transcripts, half-finished docs, dumped into a folder an AI can run around in. Hell yes.
If this is how one person solves a hard problem, what happens when the problem belongs to a team? What happens when the folder is the company?
Something changes in kind. The moment every person and every tool feeds into one folder, permissioned properly and kept current automatically, it stops being a memory and becomes raw material that future agents will absolutely devour. It becomes a goldmine. One person can change how the whole team behaves, in a single move, without calling a meeting.
Alright, now we can talk about CRM. That’s the longest I’ve ever gone without talking about CRM, thank God it’s over.
Where does CRM fit into this new problem-solving pattern? Well (we say) let’s just use the CRM as the database. Let it hold the state. Sure. Absolutely. How far does that get you?
Go back to the folder. Garbage in, garbage out is obvious to everyone the moment they think about their own problems. If I’m building a house and all I put in the folder is crayon sketches, the model will do everything it can and I will get a bad house. Opus 4.6 will apologize, Opus 4.8 will blame me. Maybe they’re both right. If I put in the town building codes, and a way to reach every supplier in my area over an API, and a way to find general contractors and their reviews and photos of their last ten jobs, wire up an MLS search tool (relax, it’s a hypothetical), a markdown graph of LEED building standards and techniques, then when I say I need this done for a hundred thousand dollars less with better energy efficiency, I get an actual answer.
Same model, same question … but a wildly different result.
The quality of the answer is limited by the resolution of the data it has to work with.
This resolution gap explains the lack of AI progress from companies you (and I) used to love.
A model can only reason over what was captured in the data substrate, and what got kept is a rusty bucket of key/value pairs. You cannot reconstruct nine months of a human relationship out of Stage:Closed Lost, Price:$4,200. AI exposes the broken data model in painfully obvious ways.
What does the model have to do when it works from a legacy CRM? It has to repeatedly sift through thin, hand-entered data and painfully reverse engineer the story ad hoc. Every time, from scratch. Why is this deal in this stage? Why is this person marked as the economic buyer? Is that even filled in? Uh oh! That sounds like it’s going to suck! Guess what? It does!
The data that is most relevant to the task in front of you … maybe the renewal plan from the CS leadership weekly meeting, or the board note from last quarter, the disposition of the lawyer on the year-making banner deal (and lawyers who approved similar deals and the reasons they did) … has nowhere to live in the CRM’s data model. Nobody would ever enter it by hand. You would need a sea of a hundred thousand 142-IQ temps who never slept. It is not realistic to get this data without AI, and in the legacy systems there is nowhere to put it. So it isn’t even a question of how fast you can retrieve it. It’s not there.
But … the models are ready. Everything I did with a folder for a school application, anyone can do for a business, around the clock, if the record is there. The bottleneck is not the model. It is what the model is physically able to access and cherry-pick into context.
The system of record is a system of amnesia. Whatever you put on top of it inherits the amnesia. Better prompting doesn’t raise the limit. Better hygiene doesn’t raise it. A rep who fills in every field perfectly is not even that helpful, because the fields cannot hold everything that actually happened. Call it the resolution ceiling: an AI’s output can never be sharper than the sharpest data it can reach, and a store built around human data entry fixes that limit at whatever a person would bother to type.
Every company that tries to run itself on AI will hit the resolution ceiling. Many will hit it without knowing what they hit. They’ll blame the model.
Some will notice and flag it. This is a fun proofpoint … check this out. My chief of staff agent researches every night, looking for the single thing in the business that would surprise me the most, and today (September 15th, 2026) it told me about a new pattern - spoiler: the pattern is what I am talking about, specifically the teams that hit the wall and know exactly what they just hit. On the discovery call the VP Sales said: “I’ve built my own skills in Claude, but it’s taking me far too long to run through every single step to build this out. Like it’s taking me 12 minutes to write, to run context and write an email that doesn’t work. I need this to scale and be quicker.” Then his CTO described the wall, saying they need to “ingest these additional net new records and incrementally build on that [context] graph” and surface signals to BDRs automatically, because if so, “we don’t have to spend a whole lot of money just building workflows and signals within Salesforce, which is building for the last generation of workflow operation as opposed to agentic.”
Let’s. Fucking. Go.
In 2025 effectively 0% of GTM teams saw or even touched this wall. Late 2026 I’d guess fewer than 10% see it, but almost all have hit it. By 2028, no GTM team will tolerate hitting it.
And that is why CRM goes away.
The original promise of CRM got abandoned by an entire industry, and I was part of that industry, and it never felt like a decision at the time. It felt like a (dope) schema. The schema is the ceiling. It is not in fact dope anymore.
The lineage is short. My grandfather was an early sales rep at IBM. Imagine being him in the 50s. You collected business cards and kept them in a stack. Imagine being a rep in the 90s. You got a computer, so you put those business cards in a spreadsheet. In the 2000s? You got the internet, so you put your spreadsheet online. Every incremental step made sense, and all along the way the pretense remained the same: the human is the source of truth. Any insight into the business, at any level of detail, comes through a person if it ever makes it onto disk at all. That was correct, because a person was the only thing on earth that could turn a conversation into data. A spreadsheet-shaped store was the right store for spreadsheet-shaped data. And someone who was reasonably good at capturing that information and reasonably good at acting on it, I think I’ll call Julie again today and see if I can get her to do the thing, was worth six figures in the enterprise and seven on a good year.
We have been switchboard operators. Back in olden times (as my kids would say) many highly-capable people inherited jobs sitting in rooms full of switchboards, physically plugging a wire into a hole so two phones could talk. That wasn’t a bad job because the people doing it were bad at it. It was a bad job because it existed only to patch a hole in the technology. A call couldn’t route itself yet, so a human stood in for the missing piece. The instant it could, the job didn’t get easier. It stopped existing. Every switchboard on earth went quiet within a decade, and nobody wrote a nostalgic essay about how satisfying it felt to connect call nineteen to call four.
Every time you open your CRM to type in what already happened on a call the software could have watched (probably did!), you are plugging a wire into a switchboard by hand, because the technology supporting your work hasn’t caught up to … the software. Certainly it is obvious as you look around “your stack” how there is a massive divide between what was built from scratch assuming frontier models existed and would improve dramatically, and what was built in olden times.
Btw, my kids also refer to TRL as existing “around the turn of the century” so, yeah. Anyway.
Being the switchboard operator CAN’T be the job. The half of the seller’s job that was switchboard, being paid to be the source of truth, goes away. All the data entry. What’s the other half? I bring wonderful news for sales reps specifically, because we still need representation. We need a human to be the UI. We trust names and faces. We connect over personal history and shared taste. I’m a buyer, and I love sales reps, and I always have. My dad was an equity analyst, and he raised me on the line that anything you ever want to know costs a nickel. Pick up the phone, call the sales line at any company, and the person who answers knows everything about the product, the industry, and the people running the place. A nickel, because that’s what the payphone cost.
The people who will be valuable in information jobs in the future will be extremely valuable, and make twice as much money, or more. There are four categories: Producers (people with good taste who can make anything they want or that the business needs), Architects (the mastermind planners who set up the foundation so the producers can run wild and not burn the company down), the Adults (who bring a voice of reason, prioritization and direction into the scene), and the Representatives (whose magnetic credibility attract, educate and persuade the target market and onlookers). Sales reps are obviously Representatives and will be worth their weight in gold.
The software engineers (Producers, Architects) have flying cars. The lawyers (Adults, Producers, Architects) have flying cars. Finance people (Adults, Architects) are starting to have flying cars.
Support goes away; the best Support folks can be Producers or Representatives, easily, and the brilliance of “outcome-based” ticket deflection apps that make noisy support requests disappear are solving the right problem in exactly and completely the wrong way. Prevent the fucking ticket guys!
Woefully, Sales and Customer Success (Representatives) do not have flying cars, and are still manually inputting stuff into dropdowns and tolerating gaps that are ridiculously avoidable.
I’m looking at you, HubSpot reps. I’m looking at you, Anthropic salespeople entering everything by hand into Salesforce.
DOESN’T THAT SEEM WEIRD TO YOU ALL? It should seem weird.
You are selling “AI CRM”. Or you are selling models that will cure cancer and you are entering the notes by hand and ALL of you are sitting through CRM hygiene trainings. You’re wondering if I know how often your team has those sessions and how often your manager hassles you to update CRM, and sadly I KNOW exactly the answers to these questions. Maybe … revolt??
Customer Success is just left for the wolves. What’s our plan for this renewal? Holy shit, we have a week. Guys. This is an open book test. All of the data exists. Models that could comb through every word of it and come back with a plan EXIST. Every decision from every management meeting and ops meeting is capturable. All of it can live in one harness in a way that makes these problems unbelievably easy to solve, or at least to make progress on that builds on the last progress. I believe go-to-market teams are very close to this epiphany. When it hits, I want there to be a credible answer waiting.
Forget the categories for a second. Let’s do a little perspective-taking …
Imagine being a customer, shouldn’t be hard since you definitely are a customer of something. Great. Close your eyes and imagine … you hit a bug, and an hour later you get an email apologizing, tells you it’s fixed (already). You mention, offhand, on a demo, that you wish the product did one more thing, and that thing exists a week later, in your pilot, because nine other prospects asked for the same thing and building it was the highest-leverage thing engineering could do that week. The cold email you get names the companies you respect and the problem you’re wrestling with this quarter, instead of telling you it noticed you went to Michigan. Go Blue! Have a minute for a quick call to hear more about LeadWeasel? (No). Btw, a quick note to the AI SDRs out there: Winchester, Massachusetts does not, in fact, have a famous cathedral. Come to Jesus my friends … are you running Llama 1.0 on a Nokia flip phone? And the support ticket that never should have existed never gets filed, because somebody screenshotted a confusing screen into Slack and it got fixed before anyone else hit it.
Or imagine being me in my current role. No one EVER has to book a one-on-one with me to bring me up to speed on anything customer-related. I have an agent whose entire job is to find the gaps in what I know and FILL them aggressively, to consider what the business is trying to do and every detail about every company we’re trying to serve, and then say: Christopher, look what THIS person just said, it matters (like the example earlier that happened today). Sam, one of our reps, comes into the office and puts his knapsack down, and we can sit quietly and think, and discuss HOW to approach a particular situation - the assumption is we both know everything about all the situations that matter at that moment in time. He doesn’t tell me that a totally new security review from an unrelated department (oo, good sign!) landed at the eleventh hour, I know about it, engineering knows about it, everybody knows about it. We decide it doesn’t need a push from us, to let it play out over the Tuesday ahead of us. Meanwhile Marketing is pulling quotes and website copy off the same calls that drive Engineering work in real time. Support tickets are just drying up entirely. It’s trippy, and it’s a lot of fun, and it means we can sit in fellowship without staring at a screen and spend the time getting honest with each other and making hard decisions. We refer to prospects and customers of the moment by their first names, we assume everyone knows everything, the debates, the levity, the product strategy, everything we do starts on top of the entirety of what has happened, and all the relevant details (and only those) are in memory in all our brains as we talk. When we need a new lens of what might be relevant, we can load that in collectively in seconds.
It’s fully fucking psychedelic, and I am FULLY here for it.
Nobody here has ever asked me for a slide or (gasp) a deck. I never show the board something that hasn’t been triple-verified; if I say we ran this many demos, that shit is audited up down and sideways, those are recorded meetings where a sales rep shared a screen. It’s as though a killer RevOps analyst had gone through every recording in Gong manually and checked. What does that cost me to pull? Zero dollars and zero minutes of anyone’s time. And there are questions I can ask that no manager has ever been able to answer …
Want to evaluate someone’s performance in a role? Show me the decisions this person made this month. Show me the promises they made to customers, and whether they kept them. Within seconds you see the quality of someone’s work on a human level, the part of the work that remains: the decisions, the promises, the trust, the judgment.
The more tightly a company is bound to the ground truth of its customers’ real experience, the faster it grows. A company bound so tightly to what is happening with its customers that the experience itself makes the company move. Small moves most of the time, occasionally a big one, always moving. That is what CRM was supposed to do from day one. We never built the machine that could do it.
Say a company launches a new product and nobody’s selling it. The reps aren’t lazy, but they tend to sell whatever already wins them deals, so they’re slow to change behavior. We might assume “more training” is the answer, or a “sales contest”. A slide deck, call reviews, managers nagging in Slack for weeks, and if you’re lucky a third of the team changes any behavior.
With a company brain, you go find every call this quarter where the new product should have come up and didn’t. Now we’re talking. You see exactly where the gap is. And you change the pre-call coaching instructions once, from your laptop, for every rep at the same time. Not a training. An edit. You preserve whatever tweaks and personalization each rep has told their agent they want. That’s a hard problem, but it’s a solved problem that will become the new table stakes. Currently Day AI is the only product that makes this trivially easy for RevOps.
A CRM company is NEVER going to do that, no matter how much “AI!!” gets screamed from the rooftops, because it’s obvious from one perspective, and totally weird from the legacy CRM perspective. This is why HubSpot called Skills Agents, and why they are still rigid and monolithic - they are ergonomic to the pricing and packaging, they fit THAT mold, not the ergonomic mold of the information worker who expects things to “just work” in the way they very obviously should.
On the data layer a CRM never actually held the muscle memory of the business. It held key/value pairs and a rough, woefully-incomplete event history. A company brain holds the reasoning, wired directly into the agents doing the work. Changing the business, in that brave new world, is just one atomic commit. Decision-made -> Behavior-changed.
So … how? Just follow the train of thought NOT from your CRM worldview, but from the “company brain” or “second brain” or “Karpathy knowedgebase” view of the world. Start from whatever folder already exists. At a more established company, RevOps has one, guaranteed. Don’t call that one a folder, no no, they call it a “repo”, because they share it through GitHub. It has scheduled jobs pulling Slack and Gong and maybe email, ways of reaching the internet, a skills directory, and they’re trying to run all of it automatically. Btw if you don’t have a RevOps team, I have news for you: YOU are RevOps. If you’re a founder looking for your first CRM, forget everything you know or think you know about CRM and start from the folder you already have, with your notes and decisions and roadmap and personas, and play the tape forward. When you do a customer interview, does it update everything? Does it change your mind? Does it reach out to you? You, the early-stage founder, can now expect the same quality of reporting, forecasting, insight and leverage that a public company with 200 people in Ops has - and actually, yours will run faster and much much better.
Think how excited you are about this “brain” view of the world, and keep that daydream going … you will hit two places where it makes sense for a vendor to fill the gap. Work backwards from your vision of the end-state nirvana and you’ll see a couple existential questions.
If the customer’s experience is going to make the company move, nothing factual can get deleted from the historical record, just to save space, or because the data substrate is too low-resolution. If you are storing “rows” you’re cooked. If you’re storing “nodes & edges” you have a chance. The legacy system of record relational database dies and a context graph is born: a living map that frontier models populate continuously, where plain language sits right beside structured data as a first-class thing, and where an agent finds what it needs by walking the graph, a hash lookup, instead of reverse-engineering a customer out of a dozen joins. Three things follow from taking that seriously. Every value carries its own receipts, because the second a model is the one filling in a field, everybody in the room wants to know why and what it read to get there; nobody ever made a rep cite sources for Closed Lost. Who may read a value is part of the value itself, decided when it is written, so the same question returns a different graph to the CEO, the CRO, and the rep, with no filter applied at read time; this is the thing a security review is actually asking about, and it is the hardest thing to retrofit onto a build that started with one god-mode API key. And it has to forget on purpose. Pull one email out and everything downstream of it gets found and destroyed, because the day a Slack channel goes private, whatever was said in it needs to disappear for the new hire and for their agent both. It is not the case, sadly, that an agent connecting dynamically to a bunch of different systems will work. It just doesn’t. For the core of what you’re working with, it needs to be in one place, and that place needs to be fast, and you cannot build it casually. It takes skinned knees and a few very talented people who at some point along the way have to be right, and then have to pull off the build in production. Salesforce still hasn’t gotten itself off Oracle onto Postgres, I heard they gave up after a decade of trying but that’s a rumor. Do your own research, draw your own conclusions.
Once everything is saved in customer memory, the data entry work ought be done by something other than a person filling out a form in an archaic “workflow app” like Salesforce or HubSpot. In fact it HAS to be done by someone other than a human because the amount of data to enter is now so enormous, and so detailed, much of it resulting from deep consideration of the entire dataset itself (ie. adding the answer/value, plus reasoning and source citations for a property titled “Biggest marketing challenge”). It is possible to have agents do the work of filling out the legacy CRM data, it’s easy to do with Day AI agents for example, or Grok Bot, or whatever. But a) most of the data is lost when written (the “why”, which really matters) b) it is generally done with wild inconsistency (a real problem, given that you can’t debug because the “why” is not saved anywhere) and c) the costs are often exorbitant; HubSpot charges $0.10 to fill a single field out automatically - this is several orders of magnitude more than it ought to cost. HubSpot has targeted Day AI customers and I heard one story where the cost of doing (a simplified version) on HubSpot what they did on Day AI would cost them $100k a year. They pay us under $10k annually.
The agentic Skill files you’re developing are probably a really good starting point. They ought to improve with use, they need to be responsive and adapt to the individual preferences and feedback from each inidividual human at the company. And you need to run them on governance rails. To do this, you run them on a real agent control plane. Your agents then take the shape of people: each one has a name, a job, skills for that job, and a person it reports to, and because it reports to one person it can improve on their feedback, match how they work, and fit their exact responsibilities, while every one of them reads and writes the same graph. Deployment that is versioned and orchestrated: you can see what every agent is doing, change the whole fleet with a plan and a diff and a rollback, and the personal layer each rep has built survives the change; strategy becomes something you can read and roll back like code. And loops that improve with use: each reply becomes the next version, skills fire on events in the graph and not only on a clock, and the coaching gets better because the rubric it writes to gets better. Solo, you get by with Claude or Codex and a cron. As soon as you want a new hire’s agent to see part of your email, or to control which transcripts a growing team can see, you have a vendor problem, full stop. The test to apply: do the agents get better over time just by being used? If not, you need a different architecture. Salesforce made orchestration its whole pitch for two full years, then spent $2bn on a product whose description is word for word, exactly the platform they already told everyone they had. The emperor has no clothes.
The need for and nature of this substrate has been obvious to anyone paying attention for the better part of a year. Foundation Capital published a piece called “Context graphs: AI’s trillion-dollar opportunity,” and it reads like a spec for what we built. We started building it a couple of years before they wrote it. In our view they were exactly right. It is hard to build, it is easy to use, it is wildly powerful, it exists and is available commercially from exactly one vendor.
There are a couple common fears, so let’s talk about them.
The first fear dominated the tech news cycle and drove the bearish SaaS sector market narrative for a weekend and has (largely) faded: won’t people just build their own CRMs? Sure. I am SURE someone will open-source something you can run yourself on Vercel or Netlify with Postgres in the cloud, and it will be as good as any incumbent, and your data will be yours and yours alone. The second you want to make it multi-player, you’re cooked.
The second fear is “we have to keep our CRM.” Have no fear. Take this agent control plane make it responsible for all CRM data entry. When people go to do some “sync” integration between the context graph and a legacy CRM, they often don’t bother. But you can, theoretically. There’s just no point. Being realistic though, it is very useful to delay moving off a system of record onto some newfangled context graph thing, and I know what a migration costs better than almost anyone alive. You don’t have to bite that off on day one. But once people are in and they see the depth of what can live in a context graph, they realize there is nowhere to put any of it back in the legacy system. The context graph is a superset. You cannot play Elden Ring on a Game Boy. And if you find yourself thinking about how you might port it, the problem is that you’re still thinking in the conveniences of a Game Boy.
Neither of these categories exists yet as a category. There is no context graph category on G2 today, despite the essay that named it. This isn’t a feature cycle you can absorb into next quarter’s roadmap. It is a full reset of a fully mature, fully consolidated industry, an Innovator’s Dilemma so extreme it reads like a parody of the original. The categories that define the next decade aren’t on anyone’s map yet. That gap is the whole opportunity.
So: what if somebody saw all of this coming early, back when the models weren’t remotely ready, and built both pieces anyway, one codebase, all pre-integrated, maybe a decade ahead of when the maturity curve says any of this should be possible? And what if you could run it right alongside everything you already have, without ripping a single thing out, and watch your existing stack get better overnight?
Roles at work combine, radically. People gain real agency at work. A prospect reads a blog post with a name and a face on it, then downloads a guide made by that same person, then gets an email from that same person and books time with that same person, buys from that person, talks to that person the whole time they use the product, renews with that person, and on the rare occasion they need support, hears back from that person. And that person is making twice as much money and maybe working eighty percent as much. They sit on a Zoom with a customer they know well and have strategic conversations. They hear about a pattern that’s working and turn it into marketing material that same day, and the leads come in, and they do it again. One of our customers is having more fun than I have ever seen anyone have with a piece of software at work.
I gave a talk at INBOUND to something like twelve thousand people. Big product releases, the whole thing. What I keep coming back to is crossing the street in the Seaport afterward, and a woman stopping me in the middle of the road. She said, I just have to tell you the story of the last year of my life. She’d been in a toxic job, commuting forty-five minutes to an hour each way, a single mom dealing with pickups and a daughter turning into a teenager. Underpaid, and more than that, underappreciated. She learned the product my friends and I built, and she put it on her cover letter. I am capable of doing the following things. Put me in, coach. She found a job fifteen minutes from her house at a small manufacturer outside Rochester. She’s present for her daughter. She makes a good bit more money. And most importantly, she likes the management team, and she walks on water to them.
That’s the bar.
CRM had a good run. I helped build some of it, and I’m proud of that period. Now for the first time in my career I get to go build the thing we’ve been after all along: fueling company growth with the voice of the customer.
That’s life after CRM. I’d rather be there than anywhere else. It may be the life for you, too.
Already have the folder? Find out where it stands. Run this from its root, then open it in Claude Code or Codex and say “evaluate this company brain.” It takes stock of what you’ve built, grades it, and writes one file: the plan to upgrade it. It’s free.
npx skills add day-ai/company-brain-evaluation