A Chatbot's Personality Just Sold for Nine Figures

Last week, AI coding company Cognition, the maker of the Devin coding agent, acquired The Interaction Company of California, the startup behind Poke. The price: low nine figures, confirmed by Poke co-founder Marvin von Hagen. Not bad for a product that launched this spring.

Poke is an AI assistant you text like a friend. It lives in iMessage, SMS, Telegram, and in some markets WhatsApp. No app, no dashboard. It handles email triage, reminders, travel, scheduling, and it talks like a person, slang and jokes included.

Here's the number that matters: over the past three months, Poke users exchanged more than 100 million messages with it. Hundreds of thousands of people were texting a chatbot enough that it became a daily habit.

And here's the detail founders should sit with. Cognition didn't buy Poke for its model. Cognition has its own models. It bought the way Poke talks to people. When a company with world-class technical infrastructure pays nine figures for tone of voice and interaction design, something has shifted in what counts as a moat.

Because the thing Cognition sells, an AI software engineer named Devin, has a problem every AI product now shares: the underlying capability is getting commoditized. Plenty of agents can write code. Fewer feel like a colleague you'd actually want on your team.

Von Hagen put the thesis plainly: "You probably prefer it if you have co-workers that have personality, rather than if you have co-workers that are just robots." Cognition co-founder Scott Wu described what they were buying the same way: an agent that's "proactive, it knows you, and it's fun to talk to. That's exactly how working with Devin should feel."

Notice what's not in either quote. Nothing about benchmarks. Nothing about model quality. The deal is a bet that as capability converges, the winner is decided at the interaction layer: how the product communicates, how it remembers you, whether people enjoy it.

You can see the same bet elsewhere. Jack Dorsey's new Buzz is a group chat product built around teams working alongside AI agents. The frontier is moving from what AI can do to how it feels to work with.

Strip away the hype and Poke made four product decisions worth studying, none of which required a research lab.

It went where people already were. Instead of another app fighting for a home-screen slot, Poke lived inside messaging apps people open 50 times a day. In June it became the first AI agent approved on Apple's Messages for Business platform, which turned a distribution hack into a defensible position.

It had one voice, consistently. Poke didn't sound like a press release one day and a customer-service script the next. Familiar, funny, a little informal, every time. People described it to friends the way you'd describe a person, and that's what word of mouth is made of.

It was proactive. Poke didn't wait to be asked. It nudged, followed up, remembered. Software that starts conversations feels alive in a way that request-response tools never do.

It picked boring, recurring jobs. Email, reminders, to-dos. Not moonshots. Tasks that come back every single day, so the habit loop had something to grip.

None of this required money. It required taste and discipline.

Now the other half of the story, because it's just as instructive.

Poke was expensive to run. Von Hagen admitted the company was struggling to turn a profit despite hundreds of thousands of users. Every one of those 100 million messages cost real compute money, and a friendly personality doesn't lower the inference bill. The product people loved was, as a standalone business, burning cash.

So be careful which lesson you take. Poke's outcome was great for its founders and investors: a nine-figure exit inside a year. But the exit happened partly because the business needed a home with cheaper models and deeper infrastructure. Being loved created the option. It didn't create the margin.

For a first-time founder, this is the whole tension of consumer AI in 2026 in one story. Delight drives usage, usage drives cost, and cost arrives before revenue does. If your product's magic depends on heavy model usage per user, you need a plan for that math from day one: cheaper models for routine turns, caching, usage tiers, or a buyer who needs what you've built. Hope is not on that list.

A fair objection: can't anyone tell their model to be funny? Sure. And most attempts land like a dad joke in a board meeting. What's defensible isn't the instruction, it's everything around it.

Consistency over months. Anyone can be quirky for one demo. Keeping a voice coherent across millions of interactions, through edge cases and angry users and weird requests, takes real editorial infrastructure: guidelines, evals for tone, someone with authority over how the product speaks.

Memory. Poke felt like a friend partly because it knew you. Accumulated context per user is a switching cost that compounds. A competitor can copy your tone in a week but can't copy what you know about your users.

Proactivity that's welcome instead of annoying. Knowing when to nudge is tuned by watching real behavior. Get it wrong and you're spam. Get it right and you're indispensable.

And fit between voice and job. Poke's casual tone worked because managing email is personal. The same voice in a tax product would be a disaster. Matching personality to context is a judgment call, and judgment doesn't ship in an API.

The system prompt is the visible 10%. The moat is the other 90%.

Here's the practical playbook, sized for a team of two to five.

Write a voice guide before you write more features. One page. How does your product greet someone? Admit an error? Deliver bad news? Decline a request? If you can't answer in your product's voice, your model answers in whichever voice it woke up with.

Read your transcripts weekly. Pick 20 real conversations and grade them against the voice guide. This is the most valuable hour a founder can spend on an AI product, and almost nobody does it.

Add memory before you add capabilities. One remembered preference ("you like aisle seats") creates more loyalty than three new features.

Ship one proactive behavior. A follow-up, a reminder, a check-in. Measure whether people respond warmly or mute it. Iterate there.

Pick a distribution surface people already live in. Email, SMS, WhatsApp, Slack.

And before any of that, get clear on who you're for and why they'd care. Mapping your users, competitors, and the job you're hired for is planning work, and structured tools like Foundra exist to walk first-time founders through exactly that exercise before the code starts.

Personality is a differentiator, not a religion. There are products where charm subtracts value.

If your users are in a hurry, in pain, or in trouble, they want competence with no garnish. Nobody wants a witty error message from their bank when a payment fails. Nobody wants banter from a legal tool at 11pm before a filing. In high-stakes contexts, the winning personality is calm, precise, and nearly invisible.

Novelty personalities also age fast. What feels fresh in month one can feel like a colleague who won't stop doing the bit by month six. The personalities that last are mild and useful, not loud.

And personality can't rescue a product that doesn't do the job. Poke worked because it actually handled email and reminders well; the voice was the layer on top of competence, not a substitute for it. Sequence matters. Make it work, make it reliable, then make it someone.

Zoom out from this one deal and there's a repeatable shape worth remembering.

Every technology wave commoditizes its core capability, and then value migrates to the layer where the technology meets people. Databases became commodities; Salesforce won the workflow on top. Models are heading the same way, and 2026 is the year acquirers started paying up for the human layer.

That's an enormous opening for small teams, because the human layer rewards taste, consistency, and attention more than capital. A two-person team cannot outspend anyone. It can out-care almost everyone.

So the takeaway isn't "add jokes to your chatbot." It's this: whatever you're building, assume the raw capability under it will be table stakes within 18 months, and ask what you're accumulating that won't be. User trust, user memory, a voice people recognize, a habit slot in someone's day. Those compound while capabilities converge.

Cognition just told the market what that layer is worth. Low nine figures, for a product younger than some of the produce in your fridge.

Was Poke's exit a success or a rescue? Both, and that's normal. The product had real love and real distribution but heavy costs; the acquirer had infrastructure and a need. Most good acquisitions are two half-solved problems finding each other.

Can I really differentiate on personality if I'm using the same models as everyone else? Yes, because personality lives in the accumulated layer around the model: voice consistency, per-user memory, proactive timing, and fit to the job. Those take months of attention, which is exactly why they're rare.

Should my B2B product have a personality? It already has one; the only question is whether you designed it. For most B2B tools the right personality is calm and precise. Designed blandness beats accidental weirdness.

How do I measure whether personality is working? Watch for users saying "it" less and the product's name more, unprompted screenshots shared socially, retention on proactive messages, and qualitative answers to "what would you miss?"

Does this mean interface startups will keep getting bought? No guarantees, but 2026's pattern is clear: companies with strong models and weak relationships are shopping for the opposite. Building something people love talking to has never had more potential acquirers.

1. Why Cognition bought Poke: AI personality is becoming a competitive advantage (TechCrunch)
2. Interaction is joining Cognition (Cognition blog)
3. Poke makes AI agents as easy as sending a text (TechCrunch)
4. Apple approves Poke as the first AI agent on its Messages for Business platform (TechCrunch)
5. Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents (TechCrunch)