{"id":90418,"date":"2026-11-13T12:21:30","date_gmt":"2026-11-13T17:21:30","guid":{"rendered":"https:\/\/paulsyng.com\/blog\/?p=90418"},"modified":"2026-11-13T12:24:22","modified_gmt":"2026-11-13T17:24:22","slug":"what-canva-can-steal-from-fin","status":"publish","type":"post","link":"https:\/\/paulsyng.com\/blog\/what-canva-can-steal-from-fin\/","title":{"rendered":"What Canva Can Steal From Fin"},"content":{"rendered":"\n
\n

“With 265 million users on a monthly basis hammering our services, we have to own our models and we have to own infrastructure that serves our models.”<\/p>\n<\/blockquote>\n\n\n\n

Cliff Obrecht, Canva’s co-founder, told Fortune<\/a> in April. Canva decided to build its own AI because renting it from other companies cost too much. He’s right about the cost. In August, Melanie Perkins told Fortune<\/a> that Canva “needed to reduce the cost of completing an AI task.”<\/p>\n\n\n\n

But look at what the statement commits Canva to. Own the models. Own the system that runs them. A company that does both has built an AI lab. So this essay asks one question. What happens if Canva takes Cliff’s statement all the way and makes the lab the company?<\/p>\n\n\n\n

One company has already done it. It used to be called Intercom, and it’s now called Fin<\/a>. I wrote this for the founders and leaders of Canva. Everyone else is welcome to read over their shoulder.<\/p>\n\n\n\n

Why Fin<\/h2>\n\n\n\n

Intercom sold help desk software. Support teams at other companies used it to answer their customers. By 2021, it was taking in about $200 million<\/a> a year.<\/p>\n\n\n\n

In late 2022, the company was in trouble. Eoghan McCabe returned as CEO that October, and by his own account<\/a> it was “heading quickly towards negative growth.” ChatGPT launched a month later. He went all in.<\/p>\n\n\n\n

Four years on, the company is called Fin, after its AI agent. Fin answers the customers itself. It solves about three in four<\/a> of the questions it gets.<\/p>\n\n\n\n

In March, McCabe said the company was taking in $400 million a year. Fin was about to pass $100 million of that, and was growing about three and a half times a year.<\/p>\n\n\n\n

Fin now runs on a model the company trained itself. Other companies can buy that model. In June, Salesforce agreed to pay $3.6 billion<\/a> for the company.<\/p>\n\n\n\n

Most people who study Fin study how it charges. I think that misses what happened.<\/p>\n\n\n\n

A software company that made a tool for people to work in became an AI lab whose models do the work. Canva is facing the same decision.<\/strong><\/p>\n\n\n\n

The two companies are different, and I’ll get to where the comparison breaks. First, the part that carries over.<\/p>\n\n\n\n

What changed <\/h2>\n\n\n\n

For thirteen years Canva sold one thing. It sold the skill of knowing what good design looks like. That skill sat in templates. A designer made a layout once, and Canva showed it to everyone who came looking. You picked one and made it yours.<\/p>\n\n\n\n

Then the skill moved. A model can now produce the first draft from a single sentence. Canva said so itself<\/a> in April: “Creation is no longer limited to a blank page or the perfect template.”<\/p>\n\n\n\n

So the skill Canva sold now lives inside AI models. And a model is only as good as two things: what it learned from, and the system built around it. Whoever owns the best design models owns what Canva used to own. That’s the problem.<\/p>\n\n\n\n

Some things haven’t moved.<\/p>\n\n\n\n

People still make things so others can see them and judge them. They still want a hand in the result. Danny Wu, who runs AI products at Canva, says<\/a> the average design “is edited dozens of times before it’s published.”<\/p>\n\n\n\n

Last year I called this voice<\/a>. Someone who couldn’t show what they meant now can. People will want that in 2030 too.<\/p>\n\n\n\n

Now look three to five years out.<\/p>\n\n\n\n

Most first drafts will start inside an assistant or an agent. It has already begun. Claude can make a deck<\/a> and send it to Canva, or skip Canva and hand over a PowerPoint. Gemini builds whole decks<\/a> inside Google Slides, where people already work.<\/p>\n\n\n\n

The main user of AI is turning out to be other software. Menlo Ventures<\/a> counted it this month. On OpenRouter, a marketplace for AI models, agents now make up 80% of the traffic. And the money follows the models. By Menlo’s count, the big AI labs now take 69 cents of every dollar companies spend putting AI to work. A year ago it was 50.<\/p>\n\n\n\n

So fewer first drafts will begin in Canva’s editor every year. Models that any assistant can call don’t have that problem.<\/p>\n\n\n\n

Play 1. Make the models the business<\/h2>\n\n\n\n

McCabe moved nearly 80% of Fin’s research and development onto the AI agent. At the time, the agent brought in less than a tenth<\/a> of the revenue. He rewrote the company’s mission around it. He ripped up the old values and wrote new ones.<\/p>\n\n\n\n

He also committed $100 million of the company’s own cash to the bet. Three years later, he went further. He put Fin’s models up for sale and wrote<\/a>: “we will now begin the processes of disrupting our agent business with our AI business.”<\/p>\n\n\n\n

AI began as a help desk feature. He rebuilt the company so that the AI was the business. Canva is one step into this. It built its own models, and Cliff has explained why<\/a>: “Building our own models gives us much more control over the cost and speed of AI.”<\/p>\n\n\n\n

That’s a reason about cost. It treats the models as a way to run the existing product for less. The steal is to treat them as the product. Canva makes models that do visual work. The editor is one place they get used. A company’s brand system is another. Other companies’ software is a third.<\/p>\n\n\n\n

All three feed the lab. More use gives the models more to learn from, and better models make all three better. The founders have already said half of this.<\/p>\n\n\n\n

Cliff, on Canva’s new image model: “It brings Canva to more people and opens up a whole new revenue opportunity for us.”<\/p>\n\n\n\n

Melanie, in August<\/a>: “We’ve essentially had to learn how to build the foundations of an AI-native company ourselves.”<\/p>\n\n\n\n

Read those two lines next to Cliff’s first one. The founders have already described the lab. They haven’t yet called it the company.<\/p>\n\n\n\n

Play 2. Build a real lab<\/h2>\n\n\n\n

In early 2023 Fin had about six AI researchers. Three years later it had about sixty<\/a>. McCabe is blunt about who they are: “real, Phd level AI scientists and researchers.” He’s equally blunt about how he treated them: “We aggressively promoted and compensated our AI team above others.”<\/p>\n\n\n\n

Inside the company, they call it the AI lab. Co-founder Des Traynor says<\/a> it’s “the most contested resource that we have.”<\/p>\n\n\n\n

Translation: every team wants the lab’s time, and its people are paid more than anyone else’s.<\/p>\n\n\n\n

Canva is ahead of where Fin was. Canva Research has more than 160<\/a> scientists and engineers. It built Craft, Canva’s image model, from scratch in about six months.<\/p>\n\n\n\n

Canva even uses the word. Craft placed fourth on a public ranking, behind three models from OpenAI. Canva’s reading, as SmartCompany reported it, was that this made it “the world’s second most advanced AI lab for image generation.”<\/p>\n\n\n\n

So the lab already exists.<\/p>\n\n\n\n

The steal is what you point it at. Images are OpenAI’s home ground, and three of its models already sit above Craft. Design is a different problem. A design is built in layers, and it has to stay on brand. Change one part and the rest has to hold.<\/p>\n\n\n\n

Canva already has a model built for that. Here is how they describe<\/a> the Design Model: “The world’s first model trained to understand the complexity of design.”<\/p>\n\n\n\n

Give the lab that problem. It’s the one Canva is best placed to win. It has the designs to learn from, and the designers who know what good looks like.<\/p>\n\n\n\n

Play 3. Train on what only you have<\/h2>\n\n\n\n

Fin’s own model is called Apex. It didn’t start from nothing. It started from an open model, the kind anyone can download. Then Fin trained it further<\/a> on years of its own customer service results.<\/p>\n\n\n\n

McCabe’s line on this is the most useful statement in the whole story: “Post-training is the hard part.” He adds: “You need proprietary sources of truth.”<\/p>\n\n\n\n

The base model is cheap, and anyone can get one. The edge comes from data nobody else has.<\/p>\n\n\n\n

By Fin’s own tests, Apex solves 73 questions in every 100. The best models from the biggest labs solve 71. And Apex runs at about a fifth of the cost.<\/p>\n\n\n\n

So what are Canva’s sources of truth?<\/p>\n\n\n\n

Start with the library. Canva holds more than 165 million<\/a> templates and assets. It trained Craft on millions of those designs, while its competitors mostly trained on photographs.<\/p>\n\n\n\n

Then the edits. 265 million people use Canva every month, and its AI tools have been used more than 38 billion times. Every time someone fixes a draft, they show what was wrong with it.<\/p>\n\n\n\n

Then the approvals. When a brand team approves a layout, it tells you what “right” means inside that company. The biggest labs have more data than anyone. They don’t have this data.<\/p>\n\n\n\n

One caveat. In 2023, it set aside $200 million<\/a> for its creators, including pay for those who agree to let their work train its models. Its contract<\/a> with enterprise customers says it won’t train on their material unless they opt in. So the lab has to earn its data. That takes longer, and it’s harder to copy.<\/p>\n\n\n\n

Play 4. Build the infrastructure around the model<\/h2>\n\n\n\n

This is the part of Fin’s story that gets overlooked. When Fin launched, it solved about 25% of the questions it got. Today it solves 76%. You might think the outside models got smarter and Fin rode along. Traynor says that’s wrong<\/a>: “very little of that is actually because of the upgrades or the bumps from the models.”<\/p>\n\n\n\n

The gain came from Fin’s own system. “Fin isn’t one thing,” he says. “It’s like 27 different things or whatever.” And every change has to pass what he calls a “torture test” of at least a thousand customer service cases.<\/p>\n\n\n\n

The model is one part of about 27. The rest feed it and check its work.<\/p>\n\n\n\n

That system is the infrastructure. It’s the second half of Cliff’s statement. Canva already has parts of that infrastructure. Melanie described the work<\/a> in August: “redesigning our whole architecture,” and sending every task to whichever model does it best.<\/p>\n\n\n\n

It has its own models. Canva says<\/a> they run “up to 7x faster and 30x cheaper than comparable frontier alternatives.”<\/p>\n\n\n\n

It has a format. Everything Canva’s AI makes “remains layered and fully editable.”<\/a> Most image models hand you a flat picture, where Canva hands you a file you can still change.<\/p>\n\n\n\n

It has a library to search<\/a> before it generates anything new. Since April, Melanie says<\/a>, the cost of an AI task has fallen by nearly 90%.<\/p>\n\n\n\n

It has a doorway for other companies’ software. Craft is available through an API<\/a>, and Canva plans to bring it to ChatGPT and Gemini.<\/p>\n\n\n\n

Fin can ask one question of every answer. Did the customer’s problem get solved? Design has no such switch. So the lab needs its own way to score a design. Did the person keep the design, or throw it away and start again? Did the brand team approve it without changes?<\/p>\n\n\n\n

I don’t know whether Canva has built this. It’s the part I’d build first, because it’s how the models get better.<\/p>\n\n\n\n

Play 5. Open the lab to other companies’ software<\/h2>\n\n\n\n

In late 2024, Fin made its agent work on competitors’ help desks. A company could run Fin<\/a> without buying Intercom’s help desk.<\/p>\n\n\n\n

In April this year it went further. It put its models up for sale, with competitors welcome. McCabe named two<\/a> in the post: “Decagon, Sierra, and the rest, you know where to reach me.”<\/p>\n\n\n\n

His reason fits in a line: “more and more value will accrue to the model layer.”<\/p>\n\n\n\n

He’s equally plain about the other side: “the differentiation that came from features and interface is, at the very least, going to diminish.”<\/p>\n\n\n\n

He thinks the app will matter less every year, and the model will matter more. So he sells the model wherever the work happens.<\/p>\n\n\n\n

Canva has started down this road. Craft has an API. Canva has an app inside ChatGPT, and by October 2025 people had used it in more than 26 million conversations<\/a>.<\/p>\n\n\n\n

Cliff’s line for it: “The goal here is to have Canva’s technology help people wherever they’re creating.”<\/p>\n\n\n\n

The steal is to go one step past people. Treat software as a customer. Imagine an agent building a sales presentation at two in the morning. It has no use for an editor. It needs something it can call that hands back a presentation, on brand and still editable.<\/p>\n\n\n\n

That’s the customer behind the 80% figure. And it’s a job Canva’s models are built for.<\/p>\n\n\n\n

McCabe also drew a line that Canva could borrow. “We’re never going to build for these specific verticals,” he wrote, “but we’d love someone else to.”<\/p>\n\n\n\n

Translation:<\/em><\/strong> let other companies build the narrow products on your models. A menu maker for restaurants, say, or a listing tool for real estate agents. They know their customers, and Canva’s models do the design.<\/p>\n\n\n\n

Play 6. Give the old product a new job<\/h2>\n\n\n\n

Fin didn’t kill the help desk. It rebuilt it, and McCabe\u00a0says,\u00a0“We’re dramatically increasing our investment in it.”<\/p>\n\n\n\n

The help desk has a new job. It’s where the support team picks up what the AI can’t finish. Canva’s editor gets the same kind of job. It becomes the place where people direct the work and sign it off.<\/p>\n\n\n\n

That job is safe. Canva’s own team has said why. Wu calls a prompt alone “an incredibly lossy and restrictive way of communicating with a system.”<\/p>\n\n\n\n

People can’t always say what they want. They know it when they see it. Then they want to change one thing.<\/p>\n\n\n\n

Melanie saw it in Canva’s first user tests, years before AI. The testers were, she said<\/a>, “really scared to click.” They got going when Canva gave them one small thing to change, like the colour of a shape.<\/p>\n\n\n\n

So keep investing in the editor. In a lab, it does two jobs at once. People finish their work there. And every fix they make shows the models what the draft got wrong.<\/p>\n\n\n\n

Play 7. Change the words last<\/h2>\n\n\n\n

Fin took its new name in May 2026. By then most of its engineers had been working on the AI for three years. McCabe’s own verdict<\/a> on the timing: “so obvious and so late, it’s almost an admission of failure.”<\/p>\n\n\n\n

So he thinks he waited too long. Look at the order anyway. The company changed first, and the name caught up. Canva has done it the other way round. In February Cliff told TechCrunch<\/a>: “We’re becoming an AI platform with a bunch of design tools.”<\/p>\n\n\n\n

I’ve argued on this blog<\/a> that a company’s position shows up in where it spends, long before it shows up in what it says. Fin is the cleanest case I know. So I wouldn’t spend a day on the label. Put the best researchers and the biggest budget in the lab. The words will follow.<\/p>\n\n\n\n

Where the comparison breaks<\/h2>\n\n\n\n

Four places.<\/p>\n\n\n\n

Customer service has a right answer. The customer’s problem got solved, or it didn’t. Design has no right answer, so Canva’s models learn from a fuzzier signal.<\/p>\n\n\n\n

Fin had nothing to lose, because it was heading for negative growth. Canva has 265 million users, 31 million of them paying, and still expects to grow about 20%<\/a> this year. That’s a stronger position, and it makes difficult decisions easier to postpone.<\/p>\n\n\n\n

Fin’s customers are companies that pay and sign a contract. Most of Canva’s users pay nothing. So Canva has to be more careful about what the lab learns from.<\/p>\n\n\n\n

Fin was sold. Canva’s founders have said they plan to take it public. So they’ll have to explain the lab to public investors while they’re still building it.<\/p>\n\n\n\n

What would prove me wrong<\/h2>\n\n\n\n

Three things.<\/p>\n\n\n\n

The biggest labs catch up on design. If a general model makes an on-brand design you can still edit, as well as Canva’s models do, the lab has no edge. Three OpenAI models already rank above Craft on images.<\/p>\n\n\n\n

The assistants shut the door. If ChatGPT and the rest stop calling outside tools for design, Canva’s models lose their way in to the first draft.<\/p>\n\n\n\n

The lab stays a cost line. If Canva’s models only make the old product cheaper, Canva gets a better margin and the same business.<\/p>\n\n\n\n

The test<\/h2>\n\n\n\n

Before Canva had its name, it had another one.<\/p>\n\n\n\n

“Before we built anything \u2013 before Canva was Canva \u2013 it was actually called Canvas Chef,” Melanie told Forbes Australia<\/a> this year. “And the idea was that you could describe an idea, and then it would just instantly appear.”<\/p>\n\n\n\n

That was the first idea. It took thirteen years and a lab to make it possible. So here’s the test I’d run. Open next year’s operating plan and find the research lab. See whether it sits under costs or under products.<\/p>\n\n\n\n

That line tells you which company you’re building.<\/p>\n\n\n\n

\n\n\n\n

Sources<\/h4>\n\n\n\n

Fin and Intercom<\/strong><\/p>\n\n\n\n