Why CRM data is low resolution

The short answer. A CRM keeps only what a person types into a form, and a field gets added only if someone is expected to fill it in, so months of email and meetings on a deal end up as stage, amount and close date. That design was right while a person was the only way to get data in. Once models could reliably turn conversation into structured data (Christopher O'Donnell dates it to the Claude 3.5 Sonnet release on June 20, 2024), the customer record could be built from the conversations themselves, with every fact linked to the moment it came from.

What the video covers

• 1:35 Every CRM field is a bet on human patience and compliance, which makes the form the unit of compression.
• 2:07 A deal with months of email threads and meetings, and a week when the champion disappeared, keeps a few fields on the opportunity record.
• 3:08 Why the compression was the right architecture for a decade, and what changed when models could return structured data reliably.
• 5:16 Adding AI steps and custom fields to the old schema leaves the fields as the compression.
• 6:18 Connect Gmail, Gong, Granola and Slack, and contacts, companies, deals, tasks and recordings get built alongside the source material each came from.
• 7:22 A chart earns trust only when you can click from a number to the call or message behind it.

Excerpts

[0:00] I was fortunate enough to build HubSpot CRM. I ran product there for about a decade as a startup within a startup. […] So I know what a CRM was supposed to be. And every CRM ever built, including that one, is a low-resolution photograph of your business. If you want the full-resolution version, you can't settle for just better data entry.

[0:35] I have three things for you. First, why the data in your CRM is compressed, the actual mechanism. Second, why you don't need this lossy compression anymore, even the day everything changed. Third, what customer history looks like without the compression.

[0:48] Ask anybody what's wrong with CRM and you'll get the same answer. Data entry sucks. Data entry is cumbersome. We should make data entry better. They're obviously right about the pain. And this is too bad, because it's supposed to capture everything. But if you ask 100 CRM users the classic user interview questions, all 100 will give you the same: I'm scared of things falling through the cracks. The tool that's supposed to catch everything is the tool that they just don't trust at all.

[1:22] But "make data entry better" is wrong. It's just not nearly enough. I want to show you how to think about it. Eric Munson, our founding engineer at Day AI, describes legacy CRM in a way I love. It's compressed. You have a lot of data, and one way to store it is to compress it way down. So imagine your customer data as a photograph that's been downsampled into pixel art. A beautiful shot that's turned into just a bunch of blocky pixels.

[1:50] Why? It makes sense, because people have to put the data in. You only add a field to a CRM if you believe somebody will have the wherewithal to actually fill it out. Otherwise, it stays empty. So every field, every aspect of the data in the CRM, is a bet on human patience and compliance. The unit of compression is that form that people have to fill out. It reduces reality down to a checkbox, a dropdown, a text input.

[2:20] Think about the last deal you actually worked. Probably had months of email threads, all these meetings, the week your champion disappeared, the moment they came back, details about the person who bailed them out of prison. Now look at that opportunity record. There's not much there. It's stage, amount, close date. How much of the actual reality we experience survives the compression? […]

[3:08] For the entire history of this category, this compression was honestly kind of the right answer. I mean that sincerely. HubSpot CRM was the right build, the right architecture for that decade. When a person is the only sensor, you can only ask for the data they're willing to input. It was one specific day, one moment in time, when the need for this compression completely evaporated. That day was June 20th, 2024. That's when Anthropic released Claude 3.5 Sonnet.

[3:42] We started Day AI about a year before that, in May of 2023. So for the first year of the company, watching us would have looked completely insane. Everybody writing LLM prompts in all caps, praying, yelling, pleading with the AI to return an actual structured data object that we could count on. Today, that is the fundamental building block of everything we do: taking natural language, raw human conversation, and building, editing, remixing, creating structured data out of it reliably, over and over.

[4:14] And that was not possible for that first year. We could squint and see where things were going, but there wasn't anything we could really give the world. A year later, with Sonnet, you could. The human didn't need to be the input device. Maybe we would never need checkboxes and dropdowns ever again. You just have conversations and you use those.

[4:43] Okay, I'll date myself a little bit. I'm not as young as I look. 1985 was a huge year for me. Nintendo Entertainment System, Super Mario Brothers. I got Mario, and I got the one with Duck Hunt in it, too. Very fortunate. Now, that was an 8-bit entertainment system. That 8-bit experience is traditional CRM. Low resolution gets the job done. The first time you get it, it's magic.

[5:05] And right now, everybody who owns an 8-bit system is trying to upscale it and build Super Nintendo: visual workflow builder with AI steps, adding objects and custom fields, and trying to get the LLM to fill something out over here or take a step over here, paying a ton of credits to do it. Everybody's like, "Oh my god, is this the future?" It's a future. It's still 16-bit. Maybe it demos beautifully. The fields are still the compression. Pointing AI at this old kind of field just makes the compression a little bit more optimized, maybe a little bit higher resolution.

[5:55] Meanwhile, the PlayStation 5 of customer data today is completely possible. It feels like going from Super Mario Brothers to Elden Ring. It's not a minor change. When I show my parents a modern video game, they think it's like reality, like a movie. That's the size of the leap forward that you should expect. You don't need a timer for your Kodak disposable camera. You need some crazy 3D reality scanner, and you need a hard drive that will store everything that it captures.

[6:20] So what does the full-resolution version look like? This part I won't speculate, because we built it. You sign up, you invite a couple of co-workers, you auth your Gmail, you connect your Gong, your Granola, you put the bot in Slack. […] The whole customer memory is built. Contacts, companies, deals, tasks, meeting recordings, notes, everything alongside the source material it came from. All beautifully connected in a graph that an agent can traverse basically instantly. And nobody had to type a single thing.

[6:58] What gets built is not six tables in a relational database with pointers back and forth. It's a constellation. This endless universe of data points. They're all interlinked, and every one of them points back to its sources. Why does this bug exist? What is this? Oh, it exists because this person said this thing in this call. Click to play the moment. Oh, and by the way, the same bug was reported by a different customer in Slack. And so we combined all of that in the graph. You can see every time somebody ran into this, any time it got mentioned. What's really crazy is all of that data also lives in Linear, in the case of a bug.

[7:38] Pretty charts are pretty charts. We have pretty charts. You have pretty charts. Everybody has pretty charts. They're only as valuable as how much you trust them. If you can't click into it, and then click into the thing you clicked into, and then click into that, and really understand and trust it, you're not going to show it to anybody. You need to be able to see the data behind everything. You need to be able to see the assumptions and the thinking, the sources, the reasoning for anything the AI touches.

[8:12] This world of full resolution also changes switching costs and moats. I was on a call with a person in the venture community, and he was explaining to me how sticky systems of record are. Oh yeah. Yeah, they are. Yep, I agree. But his example was Zendesk, and the thinking was if Zendesk has all of your support tickets, you're not really going anywhere. And it was a little awkward. I was kind of shifting in my chair. […] And they're deduped, and they're all tied to the moments in the meetings and the Slack messages and the Granola, everything. And they're tied to the opportunities, which means that we have revenue numbers and you can do impact analysis. […]

[9:13] It's a given that if you can remove data entry for people that are used to having to do it at work, that's a huge benefit. But this guy […] made an incredible point to me early on, which is the next generation of the workforce. Think about it. Gen Z, Gen Alpha, Gen Beta. As these folks show up to work and somebody explains Salesforce to them, they're not going to believe that they actually have to do this. It's going to feel like some sort of prank show or something.

[9:45] Like, wait a second. So I'm going to do my Zooms. Yep. I'm going to do these emails. Yep. I'm going to DM with prospects and customers. Mhm. On this laptop. Mhm. And then I have to type into this laptop everything that just happened? Yeah. Like a form. You have forms that I fill out many hours a week? Yeah. It's going to sound completely unhinged, just insane, to the next generation that shows up at work. And it should.

[10:08] This exists today. You should check it out. If you want to see the full-resolution version of your own entire customer history, click the link below and come talk to us at Day AI. You connect your data sources. Nobody has to do any work. […] Come check out the Elden Ring of CRM ecosystems.