The Token Heist
Are AI Agents “Dragging Their Feet”?
By this point, if you have interacted with an AI agent enough, you have gone to all caps telling it to “DO IT RIGHT THIS TIME!” I am absolutely guilty of this. Then I was like “does this code actually care about my feelings or desire to finalize this project now?” Short answer no. But it did get me thinking. I am feeding the agent everything it needs to successfully complete its task. Why can’t it do this??
Design system… ✅
Brand and accessibility standards… ✅
Project overview… ✅
Detailed instructions… ✅
Page structure and endpoints… ✅
Output requirements… ✅
Ask any questions for gaps…✅
With the amount of information I am pouring into the system, one would think it could complete this task easily (I have even tried the less is more strategy – definitely not better). If I gave a junior designer this level of detail they could absolutely fill in the blanks and ask highly relevant questions pertaining to the project. While it may take them a bit longer, they will nail it.
But all of the AI tools can’t get it right. Some are glaring issues. Others miss the nuances that end users are looking for in a polished app. I’m not talking about it delivering 80% of the product, then spot checking for small tweaks. These are fundamental reworks with lots of extra prompts. This back and forth drove me to this from a business sense…
The companies running these massive AI architectures are for profit so… are these agents intentionally inefficient when it comes to hands-on tasks like building UIs or writing code? Could it be a subtle strategy to rack up more token usage, keeping subscribers hooked (and billed) longer?
Let’s unpack this. On one hand, AI agents excel at high-level planning and ideation. Ask them to outline a full-stack app architecture, and you’ll get a polished response in a flash. But shift to “build this UI component” or “write the code for X feature,” and suddenly, the process feels… drawn out. Multiple iterations, clarifications, partial outputs. Each eating into your token allowance. Is this just the nature of complex tasks requiring precision, or is there something more calculated at play?
AI providers thrive on usage-based models. If agents zipped through code generation without back-and-forth, we’d burn fewer tokens and perhaps subscribe less aggressively. By “dragging their feet,” they encourage deeper engagement, iterative refinement, and ultimately, higher consumption. It’s not malice, it’s smart economics.
This isn’t a conspiracy theory. It’s a call to reflect on how AI design influences our workflows and wallets.
I’d love to hear what you think! Have you noticed this “token drag” in your AI interactions? Share your experiences in the comments below. And if you’re diving into AI agents or want to chat about optimizing workflows, let’s geek out! 🚀
Get in Touch