In this solo commentary, Robert Ta critiques the common corporate directive to "use AI" as a vague, board‑level mandate. He likens it to a casino where engineers are handed tokens and told to throw them at AI models, hoping for a payoff. While individual users may feel a short‑term dopamine boost when tools like Claude Code return results, the deeper question is whether those outputs become valuable, production‑ready software. Ta argues that the gamble can be mitigated by building domain‑specific harnesses and private evaluation frameworks—what Satya Nadella calls a moat. Companies such as Stripe and Replit have reportedly succeeded by establishing trusted AI tooling, custom evals, and workflow integrations that align AI output with real customer problems, thereby turning the casino into a more predictable, high‑win‑rate operation. The ultimate executive goal, he says, is to shift from a 51% house‑edge scenario to a 90% success rate where AI consistently delivers enterprise value.
Chapters
Every board, at least in America, I'm guessing around the world too, is at tech companies are like, use AI. And especially at tech companies, it's use AI to be able to deliver higher feature velocity and then capture incrementally more market share and also drive down costs to serve. And so there's this analogy that I have, which is if you're a tech executive and you have hundreds, thousands of employees, engineers, builders, they're all using AI and you have the board being like, everybody use AI. Go use it. Token Max, right? That was like a few months ago. What it is, is it's just a casino. You're telling everybody to go play, go throw in tokens at the casino, and hopefully you get something good out of it. And there's this feeling at the micro scale, which is when you're using it, you do feel productive because you get a dopamine hit every time the Claude code or whatever comes back. But there's a question there, which is, was the thing you made any good? You spent a bunch of tokens. Is this a useful PR? Does this actually land in production? And leaders like Stripe and Replit, clearly they've kind of figured out, okay, if we have domain-specific harnesses and set up our teams well, get them to trust the AI tooling, then it's kind of less of a casino because we've created the evals, the private evals, as Satya from Microsoft says, is really the moat, your private evals in your own domain. If you create those well and you get the harnesses around the problems that you're trying to solve for your customers and your own workflows to solve those problems, a.k.a. ship more software to solve customer problems as a technology org, then it becomes less of a casino. You can kind of think of it like if I'm putting in tokens and the house wins 51% of the time and I lose 49% of the time, it's like, well, how do we win like Like 90% of the time. That's sort of the question that I think every tech executive wants is, please, anytime you use Claude Code, it comes up with something valuable that increases enterprise productivity or value. I kind of like that analogy and it's been, I think that's gone over well with the customers and prospects I've talked to.
