In this interview, Muhammad Umar discusses the microeconomics of AI tokens, emphasizing that tokens are costly and should be allocated wisely. He argues that a large portion of tasks—estimated at 60‑70%—can be handled by traditional software automation without involving generative AI. Only the high‑value, "last‑mile" generative work should consume AI tokens, as AI inference is expensive and not optimized for routine tasks like code generation. Umar proposes building structured, modular systems he calls "large Legos" that perform the bulk of work via automation, with AI layers added on top to finish or refine outputs. This token‑aware architecture aims to improve AI adoption by reducing unnecessary token expenditure and leveraging AI where it truly adds value.
Now let's talk about tokens. Why the microeconomics of tokens are really important and how does it affect the AI adoption as well?
So first of all, tokens are expensive. What really microeconomics of tokens mean is like not every token is the same. Just think of like you, I mean, think about like using Einstein's mind to convert grams into kilograms, right? So it's just like, what are the tasks that you want to use for which you want to use tokens? There are the tasks that probably you don't want to use AI or tokens. So I like to say, I mean, being a software Engineer from last 15, 16 years now, there was so much latent, there is so much latent potential in plain software automation for which no AI is required. And there are a lot of gains there. Rather than ignoring those gains, you want to jumpstart into AI and just trying to build something which probably was, shouldn't be there because that thing could have been fixed using software automation. So when I talk about microeconomics of tokens, it's like, for example, if you want to use tokens for stuff that probably can be done by plain automation, then do it with the plain automation and the stuff that really needs AI. AI means, let's call it these generative AI use tokens for that. AI inherently is very expensive and it's not optimized. If you really see it's not optimized for generating high quality, let's say, software code, right? So what you want, for example, what you want is, okay, there should be a system that can build these large Legos, which is structured, modular at the first end. And then you plug these Legos into the AI system to do the last mile, to do the finishing and all that. When I say this, what it means is like do the task with software automation. That's possible through software, which I think is 60 to 70% of the work currently can be done through software automation. And the remaining of the work, you can orchestrate AI layers around it, not the other way around. So that's what it is.
