Chat Sucks. The Right Interface for AI Is Email.
Chat Sucks. The Right Interface for AI Is Email.
The chat paradigm is a holdover from the "one assistant, one conversation" mental model. But AI isn't one assistant. It's an army of workers. And armies don't take orders through a chat window. They take orders through dispatch, through structured requests that can be routed, prioritized, and executed in parallel.
Apr 26, 2026
5 min read
Welcome to Slow Ventures’ Snailmail, where we slow down and share what’s been on our partners’ and founders’ minds this week.
TL;DR:
- Context Is The 5th Primitive
- Chat Sucks. The Right Interface for AI Is Email.
- The AI Companies That Win Aren't the Ones With the Best Model
Others:
Events:Slow Security - NYC (4/28), SlopCon - NYC (5/2), Navigating The Idea Maze - SF (5/14), How Creators Win The AI Era - NYC (5/21)
PortCo Activity:Gusto (Slow II) hit 500,000 customers
Context Is The 5th Primitive
Yoni Rechtman
Context:
If you've spent any time in marketplaces, you know the four classic primitives: supply, demand, matching, and trust. These have been the building blocks of every marketplace from eBay to Uber to Airbnb. Supply gets onboarded. Demand shows up. Something matches them. Trust greases the transaction. This is the playbook and it has barely changed in two decades.
But something is shifting. With AI and agents entering the transaction layer, we need a new primitive: context. Not context in the hand-wavy "data is the new oil" sense. Context as in: the rich, structured understanding of who you are, what you need, why you need it, and what's happened before that should inform what happens next. Context is what allows an agent to act on your behalf without you holding its hand through every decision.
Market Signal:
Think about every marketplace interaction that frustrates you today. You search for a restaurant, but it doesn't know you hate loud spaces. You book a contractor, but it doesn't know your house was built in 1920 with knob-and-tube wiring. You hire a freelancer, but the brief starts from zero every time. All of these are context failures.
The marketplaces that win in the AI era will be the ones that accumulate and deploy context as a core primitive, not an afterthought. Agent-networks are the emergent network effects paradigm here. Even if a marketplace is self-driving and has no conventional interface, it functions on the same basic metal: supply, demand, matching, trust, and now context.
This is not a feature. It's a structural shift. The four-primitive marketplace was designed for humans clicking buttons. The five-primitive marketplace is designed for agents acting on behalf of humans. And the defensibility of context, the compounding of it over time, the switching cost it creates, may turn out to be more durable than any of the other four.
Takeaways:
- Context is the new lock-in. Supply and demand can be arbitraged. Trust can be bootstrapped. But deep, compounding context about a user's preferences, history, and constraints is genuinely hard to replicate.
- Agent-networks are the next NFX paradigm. When agents transact on behalf of humans, the network effect isn't just more users. It's better context per user, which makes every agent in the network smarter.
- Marketplaces without context will feel broken. We're already seeing this. The gap between "search and click" marketplaces and "anticipate and act" marketplaces will widen fast.
- For founders: If you're building a marketplace today, your context strategy is your moat strategy. How do you accumulate it? How do you make it portable enough to be useful but sticky enough to be defensible?
- For investors: The question isn't "does this marketplace have supply and demand?" anymore. It's "does this marketplace have a context advantage, and does that advantage compound?"
Chat Sucks. The Right Interface for AI Is Email.
Sam Lessin
Context:
I've been building a lot of AI stuff lately. Like, a lot. Multiple fully functioning apps and services a week. But here's the thing that's been driving me crazy: the interface. Everyone's fixated on chat as the way to interact with AI. Chat, chat, chat. And I think it's wrong.
I set up a bridge where I can email into a private domain, ask for anything, and an instance that has all my key data spins up workers to process as many tasks in parallel as I want. When it's done, it replies. That's it. No chat window. No waiting. No "is it still thinking?" anxiety. Just email in, answer back.
Market Signal:
Here's why this matters beyond my personal workflow: email is async, parallelizable, and already the lingua franca of professional life. Chat forces you into a synchronous conversation with one thread at a time. That's insane when you think about what AI is actually good at, which is doing many things simultaneously.
I'm not saying chat is dead. For exploration, brainstorming, and learning, it's great. But for getting actual work done? For dispatching tasks across your life? Email destroys it. And I think whoever builds the definitive "email-as-AI-interface" product, not as a gimmick but as a genuine workflow, has something really interesting.
The proof is in the numbers: this was the highest-engagement thing I posted all week. 127 bookmarks. People aren't just liking it, they're saving it, because it's a genuine "why didn't I think of that?" moment.
Takeaways:
- Chat is a synchronous prison. AI's real power is parallelism. Chat forces serial interaction. Email doesn't.
- Email is already structured dispatch. Subject lines are task labels. Threads are project contexts. CCs are stakeholders. We've been training on this protocol for 30 years.
- The "right" AI interface may not be new at all. Sometimes the best interface is the one 4 billion people already use every day.
- For builders: If you're designing AI workflows, ask yourself whether you're building for conversation or for dispatch. The answer changes everything.
The AI Companies That Win Aren't the Ones With the Best Model
Will Quist
Context:
The Cursor/xAI deal clarified something we've been thinking about for a while.
The model rat race still matters. But moving up the stack—models → workflows → outcomes—pulls you out of it. Owning outcomes should mean higher LTV and better margins than selling model access.
Market Signal:
The companies that win won't be the ones with the best model today. They'll be the ones who can monetize the current model fast enough to fund the next training run while simultaneously developing products that don't depend on model leadership.
That's a two-variable problem. Almost everyone is solving for one.
Pure model companies are in a brutal race where the frontier moves every few months. Pure application companies are model-agnostic, which means their competitor can swap to the same model tomorrow. Neither position is durable.
The durable position is in between: proprietary data or workflows that make your model better, combined with products that capture value regardless of which model is best on any given Tuesday.
This is what the Cursor deal is actually about. Elon isn't buying a code editor. He's buying an option on a data asset (developer traces) that could make his model competitive in the most valuable AI category. If it works, $60B. If it doesn't, $10B for the experiment. That's not a valuation. It's an option price.
For anyone curious how it plays out: there is 0% chance Elon is writing anyone a $10 billion check.
Takeaways:
- The model alone isn't the moat. Compounding data assets are: traces, context, proprietary workflows. Things you can't replicate by swapping an API.
- The right question for founders: are you building on top of a model, or building something that feeds back into one? The directionality matters.
- The right question for investors: what's priced in? If the bull case requires model leadership for three years, you're probably overpaying.