Notion's Ivan Zhao launches Custom Agents — autonomous AI teammates running background workflows without a Mac Mini

uh, this is hilarious image, but uh, we will return to the timeline after our next guest because we have Ivan from Notion in the Restream waiting room. Welcome to the show, Ivan. How are you doing?

Hello, guys.

Good to see you. Good to have you on the show.

Long overdue.

Long overdue. We're so excited.

Um,

first time we've ever written.

Fantastic. Uh well, we're glad we caught you on today because there's a big launch in Notion World, but I'd love you to take us through it. What was announced today?

We're launching customer agent today. It's one of the first, if not the first, multiplayer agent product for knowledge work.

Oh,

so it does real work for you in the background. Very easy to set up. Host in the cloud. Yeah.

Uh connect to all your work products. And the best part is you don't need a Mac Mini.

That's a good line. They are going out of stock. I don't know if you've seen um but the the Mac Mini is in short supply. So uh walk me through some of like the most obvious use cases like notion is I I think of the amazing because you have what is essentially a document but also a spreadsheet and you can kind of move between different data structures and visualizations on top of data in a sort of consumer app uh UI is. And so I could imagine uh creating a document and then having an agent go and do a bunch of work to populate extra fields. So where are you seeing uh or where are you excited about these agents actually taking hold in the product?

So what you're describing was the notion probably two years ago.

Yeah.

Two years ago like during the SAS era our strategy has been consolidating all different use cases into one product. Okay.

We about knowledge base you're talking about documents.

Yeah.

Talk about project management. Yeah. And we want to we have been bringing together into one tool that's very flexible. Okay. Like for example um ramp actually the the companies that sponsor you guys. Yeah. We bring ramp last year as a new customer for notion. Okay.

And we helped RAM consolidate half a dozen different tools that the core collaboration stack onto notion. Yeah. So they don't have to pay as much money for all the tools. That number one their team don't have to jump between those tools. That's number two. I would say the best part is now they have one place to do their core collaboration work. They have one place to deploy AI.

So now notion is the core agent orchestration layer for ramp.

Um

the the product we just launched today customer agent ramp has been an early customer for us for a couple months. Yeah.

Um they're running all the enable sales enabled process a lot of internal bug triage all different process on this. But because they have people have one place to work, collaboration, system, record, truth

and one place to do their busy work, delegate busy work too. Yeah.

So emp model is what is it? Uh money and time save both and this is we're doing for ramp at the moment.

I love it. Uh yeah, talk to me about uh the the agentic cron job that feels like something that we're starting to taste with open claw. There's clearly demand for it. uh it requires a little bit more upfront effort than just firing off a deep research report or saying, "Hey, hydrate this text. Expand, contract, expand, turn it into bullet points, turn it into paragraphs by back and forth all day long." Um, but I feel like the for most businesses having an agent that's effectively on a cron job, maybe you don't call it a cron job, but it's something that runs every day, that runs over a knowledge base, over a customer list, over documents, and and does the things that AI is great at every day. That feels like something that could be incredibly powerful. How are you thinking about longunning agents, cron job agents, scheduled agents? Yeah, crown job is a pretty good word for it. Like a lot of knowledge war is kind of just crown job. Yeah. Right. So you update your uh pushing paper back and forth and crown job to from this person to the other person.

I think the the world sort of taste this power of when agent connect with the crown job through product like uh open clock. It can do a lot of work for you and so you no longer have to prompt it. It trigger work on the background autonomously asynchronously for you. Mhm.

Um our our interesting less about open cloud or Mac minis is what does this do for real business.

Yeah.

And real business is you require enterprise grade permission. It has to be multiplayer. No longer just for a personal tinkerer with your own back meaning. You have to power the entire teams with it right and it has to be easy to set up. So you don't have to be an AI tinker or AI engineer to do it. Uh you have to have the state-of-the-art models usually the day release. So all the service will provide for businesses to take the spirit of bronop background agent clock you might say uh open call you might say into businesses that's the positioning of this product.

Sure. So uh talk to me about where the capability frontier on the agent side is today. Uh I mean because agents can be turned really loose. You can give them access to Python and they can talk to any API. They can write their own CLIs at this point. And so uh you mentioned like no Mac mini but is there a world where I tell like just for our example like uh I want a new notion document generated every day with a breakdown. Uh I want you to go to a readonly access API for the YouTube API. Pull all of our analytics pull all the chat feed synthesize all that and put together a notion document that I can review with the team in the morning that says oh this segment of the show was particularly great. Here's how the analytics changed. Here's where the viewer spikes were. All of that that would require talking to an API. What does that look like if there's not an off-the-shelf integration?

Um, all these is should be possible if it's not really possible. Getting the YouTube API, getting the transcript. I don't know everybody have access to it. Gemini might have a special access to that, but assume video transcript. Um,

all this is possible because all you need to do is a runtime that can run model.

Yeah. All you do is a runtime that can talk to external APIs through code that written by models.

Yeah.

And and uh a model that's does the crunch job periodically based on certain triggers. Yeah. I just described those core ingredients, but they're the core basically the core ingredients for notion custom agents. Sure. So you not only can do those can connect to your emails, connect connect to your Slack if you guys use Slack and send your message every morning. So you don't actually have to come to notion to see the work been done. You can stay where you are today. Okay.

How how how have you processed the last uh couple years of vibe coding because when I uh the first company I ever started or first not necessarily the first company but the first like real business we started on notion and at the time uh this company does like a bunch of um it's like a ad network on YouTube and so we had a bunch of different like ad buys happening. we needed to be managing that process with the client as well as the creators and so the entire company from the beginning ran on notion and I looked at every possible SAS solution at the time uh but I looked at all them I would have needed to a lot of them didn't even for like you know work with customization so I just built all these dashboards that helped uh that helped kind of like manage all those different processes and that already had collaboration built in that already had like the account functionality. So it just like worked completely out of the box. So in some ways at that time I was already replacing like vertical specific software with notion. And so in some ways like I feel like this whole process and explosion of people being able to create different applications for different use cases is kind of like just a continuum from notion's inception. But

yeah, we started we were never a lot of people think Notion is um document tool, collaboration tool, note-taking app, relational database tool. That's never been the intent. Notion started as a computing tool. Like I really care about okay, I'm a programmer. The power of computing is in the hands of you, the programmers. How do we open up to more people? That's why the company started.

So the spirit is always has been consolidating the the fragmentation of SAS for the past five plus years. And it turns out that strategy works quite well with AI because once you consolidate those things, you have one contact to power the language models, right? That's one. Number two, because we've been taking a stance that we don't want to inject our opinion how you should run your business. You should we should just provide the Lego blocks and you can decide however you want to run those Lego blocks. So we haven't been hardcode those business logic into our apps. So and back then there's a bug what we call no code, right? And in some people talk about SAS versus the language model. A lot of SAS is hardcode the logic into your vertical apps and we don't do used to be a weakness of our product because it's how open-ended it is. Has to be require some technical minded people to use it. Turns out to be a strength because now language model can use those notion building block to do a lot of work for them. So now we're with this new product we're launching is not just working with information in and out of notion. It can power agent to work with external tools and do those cron job to do those repetitive knowledge work. So do those busy work for the company. Uh internally we call this like let AI do the night shift. So you can do the day shift.

AI can do the night shift. We to go a little bit dark mode this time because truly it's doing the night shift for us.

And nobody wants the night shift. I like the day shift. I did the night shift back in college. Yeah.

It's not a it's not a fun.

That's great. Uh yes. So, so you mentioned Gemini, thank you. Another another TVPN sponsor. Uh, but I imagine that you're pretty model agnostic. Uh, I'm interested to know how you're thinking about the different uh LLMs and then also how much do you want to surface to the user? Like I was talking to Salesforce's Slackbot and it wasn't upfront with me about exactly which model was under the hood. Now I'm a nerd and I'll ask okay is it 3.5 or 4.6 or 5.2 too and I'll have some opinion whether or not that matters who knows but uh do you want to have model switchers model pickers do you want to be at that level of like empowering the user to pick the right tool for the job or do you want to handle that internally

we do both so if you're like a normie or normie plus+ using notion yeah you can just use notion without pick the model you be auto version right um but you have more sophisticated creative custom agent that do work triage work for you you different model have different strengths and weaknesses

so you should be able to pick the model

for our strategy and I think for a lot of non-labs it's very important to be model agnostic

the labs going to get better and better do model going to do more and more but one important strategic point is uh labs don't work well with other labs models

so there's importance position to be the Switzerland of agents Switzerland of the models and that's the position we're being With the product launch, you can work with clock code out of the box, can work with cursors agent out of the box and you pretty much can pick any models uh you want that's state-of- usually the day that those model are released and as a user of those product you don't have to worry about that.

Uh I want to revisit this Wall Street Journal article that you were featured in uh back in August of last year. So, uh, the quote was, uh, Ivan, the CEO of Notion, says that two years ago his business had margins of around 90% typical of cloud-based software companies. Now, around 10 percentage points of that profit go to the AI companies that underpin Notion's latest offerings. How has that changed? Is it still 10%? Is it climbing? Is it falling? What are your predictions for where that goes?

It's not as far as 10, but it's definitely meaningful amount. Um before you you can do pure SAS margins now model people all our product are powered by AI now

majority of product powered by AI now you have to uh model provider have to take some of the margin and we're okay with that um I think we see the market change on both fronts first we we want to use state of our most capable most intelligent model because our customer wants that they want to ease the customer they don't have to worry about that and second there's a new wave of open-source foreign models are And that's why we have to be model agnostic and we can shift to different model for different type of work and that will help us with the margins and at the end of the day

our customers haven't to worry about this.

What we provide is less about model capability has been there for almost for a year or two years to do a lot of knowledge work. Yeah. What's missing in the market is this infrastructure layer that glue together model capability, glue together permissions and to provide real knowledge work for the customers at the same time backward compatible uh having a good UI for yeah company of sizes.

Uh I have one last question we'll let you go. Um, how are you thinking about sort of like it's crazy to call them legacy AI workflows because they were probably implemented like a year ago, but uh when I just think about like document summarization or even like spellchecking grammar like that was probably moved to an LLM that was capable a GPT4 class model can do that at a very low cost and maybe you want to optimize that even further by going to an open-source model on commod oddity hardware really drive down the token cost. Have you left any AI workflows in place on legacy models or have you migrated everything to the frontier and you're just moving with the frontier?

We're moving with the frontier by and large because that's what customer want. They want smarter things.

Yeah.

But things like avoid dictation, summarization, legacy model can do that.

Sure. Right.

And so it's just getting

I would say the most important part is like the market is changing so fast right now like that nobody know where the future holds but we know the model capability gets better better. Yeah.

And we always care about building beautiful and powerful tools and AI is that tool today. How do we make sure that our company can benefit from this? You don't have to be Fortune 500 to for deploy engineers. You don't have to be a San Francisco startup to have AI engineer on your team to use this. Right? Every business can benefit this technology. That's our ethos and that's why we're building this product to make it super simple. You don't have to worry about Mac Mini, not worry about models.

That's the tagline. Is that on the homepage yet? Don't don't worry about Mac Mini. We got you.

I think that's too specific calling out other products, but Night Shift is a good one.

Night shift works. I love it. Someone in the chat, John John Palmer in the chat was saying, "Well, but if I can't if I don't have to use a Mac Mini, what will I spend all my time setting up?"

Yeah.

Like to tinker. Tinker is sometimes more than 50% of the fun of it besides being productive.

No, this is true. This is true. People People want to tinker. They want to play. They want to explore and have fun. And uh it seems like it's a great time to be uh running Notion. Like it's just a very exciting time. There's so many new products you can build uh so much faster than ever before. So, congrats on all the progress.

Yeah, congrats to the team on the launch.

Thank you so much. Great to finally have you on.

We'll talk to you soon.

Cheers.

Let me tell you about Labelbox. Reinforcement learning environments, voice, robotics, evals, and expert human data. Labelbox is the data factory behind the world's leading AI teams. And let me also tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. I'm not going to leak