“I love my CRM” is a sentence I really never thought I’d hear a human say.

CRMs are classic overcrowded software. Lots of cooks have done “B2B SaaS CRM”. Hundreds of vendors selling the same grid of contacts, all inspiring the same emotion in their users: nothing. Software you’re forced to use, not software you choose.

But when I first wrote about how Attio grows, my interest got peaked for a simple reason: I’m a Notion fan, and I kept hearing Attio (a challenger CRM to Salesforce) was playing ball the same way—flexible systems, LEGO-like building blocks. Designed really well for builders who want to shape their tools instead of being shaped by them.

I still use Attio to run my entire newsletter sponsorship business, and I’ll say the sentence I thought impossible…I actually love my CRM.

My use case is pretty basic compared to how companies like Granola or Wispr Flow run on Attio (I’m a one-man shop selling ads, not a GTM team running pipeline). But the moment I knew this thing was different came within minutes of onboarding.

I connected my Gmail and calendar, and they just built my CRM for me. Quick sync read my emails and built my business: the sponsors, the contacts, the companies, my actual ICP. Records I didn’t create. Nothing imported and nothing typed in.

But this isn’t about Attio, it’s about a genuinely useful product play worth unpacking that they happen to be running super well.

Because at no point did I ask for any of that. There was need to find a chat window; hunting for a co-pilot to say “Hey Attio, please set up my CRM.” No prompt engineering.

It was an AI-shaped outcome by a product before I even knew to want it.

Meanwhile in Kansas, the lion share of products that are “AI companies” now just have the same feature: the sidebar/copilot/chat that doesn’t actually do much without you doing much.

Simply, some of the best AI features run before anyone asks it to. Ambiently.

p.s real quick guys — I’m chatting to a bunch of product and engineering folks at Cursor tomorrow and plan on making my first video (fingers crossed)…if you have any questions or things you’d want to know (think tips, roadmap, whatever)…reply here, let me know, and I’ll ask and report back!

A good friend of mine recently went full time consulting and has started posting on LinkedIn, doing socials, etc. He even wants to start hosting events in NYC for growth marketers. I said to him “Dude, trust me, start a newsletter”. And I meant it.

A newsletter is how to build an audience you can reach, and there’s people with tiny newsletters using it to reach new VCs, send investor updates, drive top of funnel to high margin products…in his case, create community for events and leads for consulting.

I also said today, beehiiv is the one place to build if you’re just starting. It’s free up to 2,500 subs, has the most creator-first tools you build/grow/monetize.

If you want all the features like I do…use my code THEDIFF30.

🥷 Instead of shipping AI your users have to summon to be useful, put it in the data layer, for example as fields that fill themselves. The goal is useful work gets done before anyone thinks to want it.

If you had to sniff a random companies AI roadmap right now, it would probably have a familiar smell: a chatbot bolted onto the product. Likely in the sidebar. Very likely an empty text box that says “Ask me anything,” which in practice means “do some free labor to discover if I’m useful.”

I just did the sniff test on my own travel roadmap, and I can confirm, it does smell that way.

I’m now thinking how to do what Attio did, who made the opposite bet. I see a similar pattern in several other big companies. Their AI lives in the data model.

Their AI attributes are custom fields on any record that generate their own values. You write what you want to know once, as a prompt on the field, and it runs against every record automatically. Including new ones as they arrive.

A good example is the “ICP fit” field I use. Instead of a dropdown someone (me) fills by hand (nobody fills it, let’s be honest), the field carries a full evaluation rubric: “Tier 1: B2B SaaS, 50–500 employees, raised $10M+, not a competitor…” And Attio’s research agent classifies every company against it: reading the record, researching the live web, and returning an answer with a confidence rating and citations.

For a one-man media shop, that’s a job I used to do by hand which now runs as a column while I do other more fun and interesting things.

Same thing is happening with their new Call Intelligence feature. Sales calls get recorded, transcribed, and the qualification fields auto complete from the chat. The call just goes right to structured data on the right record vs having to ask AI to do it or God forbid…do it by hand!

Notice what’s missing from all of this.

You and me.

No prompt box. No asking. No “discombobulating”.

The idea is something that does the homework for the user before they have to do anything.

Let’s go deeper, but first a very quick word from Granola.

Conveniently for today, Granola is actually a great example of the consumer proof of this idea. They’re an AI notepad with no bot that joins your calls, wherever you take your meetings.

I never press record. I never ask it for anything mid-meeting. It even helps me prep on autopilot for other calls. I often forget it’s there until it tells me what I need to know or would have forgotten later. That’s ambient AI doing its thing: the work happens whether I remember the feature exists or not.

If your meetings still involve a bot awkwardly announcing itself to everyone, you can fix that before your next call.

We’ve done the copilot experiment now as an industry, and the graveyard numbers are worth looking at. Only ~5% of companies that tried Microsoft Copilot moved to real deployment, and data suggest only 20–30% of licensed seats get used even weekly. Microsoft also retreated their sales targets for AI agents by 50%.

This isn’t at all because the models are bad. They’re great. One hunch is because of where they live in the UX. A sidebar sits next to the workflow, not inside it, and it asks the user to stop, context-switch, and think of a prompt. Work is friction and anything that makes someone think will 200% cause drop off. I’ve seen this in my own apps with the questions I ask people.

The moment your AI feature requires someone to think abstractly about what to ask, usage collapses.

And the other fight you’ll probably lose is that the sidebar, when it beats user inertia, has to compete with Claude.

beehiiv has an in-editor AI for instance. I love beehiiv. But I never touch it. When I want AI help, I go to Claude, because that’s where my context, my projects and my habits already live. And now that beehiiv ships an MCP, even when I want AI working on my newsletter ops like segments or reporting…I’m still in Claude, driving beehiiv like the Wizard of Oz from there.

Which is the pothole for every in-product chat: all of your users already pay for an assistant they like better than yours, and the products they love most are giving them access via MCP anyway. In-product AI is a hard thing to win.

But a column never enters that fight. It’s not asking anyone to choose where to chat. It just runs on your data, inside your product, before Claude or anyone else gets asked anything.

Think of Chat AI as pull, and Ambient AI as push. And the problem with pull AI is people have to remember it exists, believe you’ll be useful, and spend effort finding out. Over and over again.

Now as we all know in-product AI has grown past plain chat. Again, beehiiv’s assistant will actually do things like website builder, write, translate and tune content right in the editor if I asked it too. There are products doing real work from a prompt. So the line worth drawing is not as simple as chat vs no chat.

What matters most is where and how the output lands.

When the answer lands in a conversation, it gets read once (maybe) and then it’s gone. Usually disconnected from the thing it was about.

When the answer lands in a field though, it’s structured data. An “ICP fit: Tier 1” value can be filtered, sorted, counted and, the big one, used as a trigger. In Attio, an AI attribute feeds directly into workflows: new record arrives, research agent classifies it, Tier 1 companies drop into a sequence, Tier 3 may not hear from me. In this case, Attio’s AI judgment becomes infrastructure other things run against.

Leave one-off answers for ChatGPT. You go build for structure that compounds.

A copilot provides zero value until someone talks to it so it’s value is flat no matter how much the customer grows.

Ambient AI scales with the data. Every new contact that syncs in arrives pre-researched. Every call ends pre-logged. The more a customer uses the product, the more visible work the AI has done, without a single extra action from anyone. Your product shows itself off every morning when the user opens it and finds useful stuff done.

This is a great way to build those ROI receipts you can show your customers that do the work that your retention metric loves. e.g “Morning Jaryd, we found 10 new contacts for you last night, 3 are a top tier fit for sponsorship, and we’ve scheduled an email to them this morning at 9am. Have a good one!”

I’m not saying kill every chat interface. Companies that are doing ambient AI are not not doing the sidebar. You’ll still find one in every example you see below.

Chat is good for exploration. The open-ended questions you’ll ask once, in weird ways. e.g “Compare my Q2 sponsors by market and tell me something surprising”.

But in my humble opinion, having AI chat does not make you an AI company no matter how much you change the copy on your website to say so. Ambient AI by itself doesn’t necessarily either, but it certainly gets you much closer to that.

In the Attio example, their investment in their Universal Context + AI primitives +various ambient AI features is what I think positions them as an AI CRM platform vs a CRM with AI.

A handy rule of thumb → if the same question is likely to be asked multiple times, make it ambient. If it’s a one-off question, keep it chat.

Here are a few different examples each with their own lessons

There’s a good story here with Microsoft Copilot. Same company, but with the other Copilot. The GitHub one. They’re doing inline completions which now have 4.7 million paid subs (+75% YoY) and write 46% of the code their users ship. Exact same models and even same parent brand with very different outcomes. The difference is all placement. The one that’s clawing back is the side panel, the other ramping up is in the ambient keystroke path, unprompted, and costs one tap to accept or zero to ignore.

I’m seeing the same super useful trend in the ecosystem of IDEs and coding agents I build my apps in. Cursor and Claude Code have started running checks and validations in the background without me asking for them. The quality of my apps has gotten way better because of it. I used to have to define the testing and QA, now it’s ambient.

Takeaway: put the AI’s output where the work already happens and make accepting it one gesture (or auto). Less navigations—more “default on” modes.

See inline completions in action →

Finance isn’t really where people want a creative chatbot who might make mistakes. So Ramp launched the opposite: AI that codes every expense, matches every receipt, and flags policy issues as the card swipe clears. Nobody asks it to. The result of making AI invisible instead of chatty: $1B in annualized revenue and a $32B valuation by late 2025.

Takeaway: in high-stakes domains, ambient AI wins because it builds trust through visible finished work.

Read “Let the robots chase receipts” →

Radar has been running today’s play since 2016. Every single payment that touches Stripe gets scored for fraud by a model reading 1,000+ signals, verdict delivered in under 100 milliseconds, correct essentially always on legit payments. Most couldn’t tell you it’s happening and just trust it works. It cuts fraud by 38% on average and it’s a big part of why people are fine with Stripe’s take rate.

Takeaway: the best infrastructure companies have been shipping ambient ML. What LLMs changed is the type of question a column can answer: fraud scores were math, “is this my ICP?” is judgment. The judgment problems just opened up.

See how Radar decides in under 100ms →

Every AI notetaker before Granola shipped the same design (I know because I used them). A bot that joins the call, requires being let in, and makes everyone tighten up for the transcript someone might review. Granola deleted that ask and just works in the background. Same models as everyone else. Completely different product, purely because of where the AI sits.

Takeaway: the differentiation isn’t the model—everyone has the same models. It’s whether the user has to ask.

See (an try) the no-bot meeting assistant →

p.s Granola runs their own GTM on Attio. Ambient AI game see Ambient AI game.

TLDR: Push don’t pull. There’s a lot of ways that can go, but to do it like Attio…

Power move: next time someone at work says “we should add AI,” ask them one question: will the user have to ask for it?

If yes…then probably no.

› My OG deep dive. How Attio grows, their origin story, a breakdown of their PLG iceberg, and more. → How Attio Grows: The Playbook For Disrupting a $280B Giant (35 min)

› The primary. Attio’s docs on AI attributes: prompts on fields, bulk runs, auto-triggers. → AI attributes (6 min)

› The strategy behind it. Attio’s own essay on why AI belongs in the data model and not bolted on top. → AI and the next generation of CRM (8 min)

› The cautionary tale. What the copilot experiment actually did at enterprise level and why “next to the workflow” is where AI features go to die. → Copilot adoption: real usage vs hype (7 min)

🛠️ Tendem by Toloka: AI agents with real human experts in the loop. Brief it like a teammate, get finished work back. All in Claude/GPT → Try Tendem (*ad)

📺 A video I loved — “Our Minds Are Weirder than You Think.” What exactly is a mind, how and why did it evolve, and could it be the most secret place in the universe? A gorgeous 20 minutes for the part of you that ships products to other minds all day. → Watch how weird our minds really are

📬 A read I enjoyed — Collaboration sucks….“If you want to go fast, go alone; if you want to go far, go together” This phrase will slowly kill your company and I’m here to prove it. → See why “ Collaboration sucks”