For the past two months, I have been trying to sell tokens to both consumers and businesses. Some lessons that cost 300K :)
The same question kept coming back, in different forms: who actually needs them?
Consumers want a problem solved
I built two products and tried to sell them on X.
The first is a multi-agent workspace where agents can talk to each other: your Claude chat can talk directly to your Codex chat. The second is a token wrapper, a desktop app similar to Codex desktop, but with a choice of models. A desktop interface makes it much easier to use.
The second did better. The first did worse.
People don’t need fancy AI products that are fun to play with. They need their problem solved. Model selection wasn’t a factor either. We offered everything from the best models to the cheapest.
For growth, we bought blog posts from KOLs and views on X through the official advertising channels. We ran two KOL campaigns, with product upgrades in between. The data was amazingly consistent: about $6 per signup from KOLs, about $5 from X ads.
Conversion was worse: one out of 10 or 20 signups paid. A paid customer could cost $60 or more to acquire, and we couldn’t see a way to earn that money back.
I checked afterwards, and our numbers didn’t seem unusual: one analysis puts the paying share of consumer AI users at about 3%, at $5 to $20 a month. That made me think the problem was the product category, not just the marketing.
We also did customer discovery calls with paying users. Why exactly were they using this? How did they want it improved?
Only three answers kept coming up: writing scripts for social media, vibe coding, and research. That’s it. We weren’t alone in seeing the pull of coding: it now accounts for more than half of OpenRouter traffic.
From where we sit, the market for tokens feels like it is cooling. People are less excited. They’ve realized: okay, this is cool, this is amazing, it does something impressive in no time.
Then comes the next question: so what? How do we make money from here?
Businesses don’t know where to begin
The business side is even more direct.
Through friends of friends, I spent a long time working closely with three people to see where we could create value: one in construction, a very traditional industry; one who owns a chain of supermarkets; and one doing consulting that connects Chinese businesses and governments with the Middle East. There is also my co-founder, who owns a health-tech company.
Their educational backgrounds vary, including a Cornell master’s degree. None has a CS background.
The central problem is simple: most of them don’t know why they need tokens.
My co-founder is the exception. He uses tokens all the time. His company has turned its software development process into an AI harness, and all of its code is now generated by AI. They do see efficiency improvements, but there is a catch.
Any change, however small, takes a minimum amount of time because it has to go through the pipeline. A human proposes a request and discusses it with the AI. The AI proposes changes. Then come review, coding, testing, and checking that everything works.
The human still initiates every change. If they skip the pipeline, the AI won’t be able to digest that change later. In their workflow, changing one line of code can mean waiting a day, sometimes two.
The other three tried our product and the best models. Their difficulty came earlier: they didn’t know how to use AI, or couldn’t think of a way to use it.
One of them wanted to delete a file. I said, you can ask the AI to delete it. As someone not very familiar with computers, the idea that he could ask AI to do that had never occurred to him. Transforming his business with AI to create more value or reduce costs was much further away.
Even reducing costs has its own friction. When letting someone go involves N+1 or N+2 compensation, the immediate cash-flow burden comes before the savings.
This gap isn’t limited to my friends. One US analysis reports that fewer than 20% of companies with under 20 employees use AI in their operations, compared with 37% of companies with 250 or more.
The gap is where the work is
We are kind of stuck.
On the model side, I personally don’t sense the improvement anymore in the newest versions of GPT and Fable. On the application and industry side, I see a large, large gap. I believe this is where the opportunities are, but the gap is huge, and so is the effort needed.
This is only one side of the story: my own perspective. I’m not sure how the token market is going globally. But I’m pretty sure what I’m seeing is real.
A few months ago, before I left the Bay Area, my acupuncturist told me the same thing. She was attending all the courses on how to use AI and was still lost about what, exactly, she could do with it. After these longer conversations with business owners, her confusion feels much more concrete to me.
AI has to live inside specific scenarios and use cases. Figuring out how to use it is not trivial. For it to work, it needs to be embedded in something people already understand how to use.
Andrew Ng has compared AI to electricity. The way I think about it is this: people don’t need to understand electricity to know how to use a light.
You need to productize it, make it easy to use, and solve a problem people actually have.