As AI usage grows, why might it not all go to the most expensive models?
On October 9, Microsoft introduced Decision 1, designed specifically to handle classification, routing, and workflow decisions. It returns a result and a confidence score from a set of predefined options; its focus isn’t on generating a long response.
This affects which provider gets paid at each step of a task.
For example, a customer service system might first determine whether an inquiry is about a refund, shipping, or a technical issue, then route it to the appropriate workflow. Classifying an inquiry and interpreting a complex contract require different capabilities, output lengths, and costs. There’s no need to assume that every step must use an equally expensive general-purpose reasoning model.
My view is that businesses will increasingly prioritize dividing AI work by task. Simple decisions can go to suitable specialized models, while complex problems can be escalated for further handling. What really matters is comparing the cost and error rate across the entire workflow—not just how cheap a single model appears to be.
This also explains why usage volume doesn’t directly translate into revenue: as the number of calls grows, the mix of tasks and the price per call may change. If errors lead to repeated work, those apparent savings can disappear.
When looking at AI-related assets such as TAO, FET, and RENDER, it’s also worth asking what paid demand each one actually has. Microsoft’s product launch is not evidence that these tokens are generating revenue.
I’m more interested in where businesses are spending their budgets—and whether that step is genuinely worth paying for.
$TAO $FET $RENDER #AI
Tap my profile picture to view my live trades
On October 9, Microsoft introduced Decision 1, designed specifically to handle classification, routing, and workflow decisions. It returns a result and a confidence score from a set of predefined options; its focus isn’t on generating a long response.
This affects which provider gets paid at each step of a task.
For example, a customer service system might first determine whether an inquiry is about a refund, shipping, or a technical issue, then route it to the appropriate workflow. Classifying an inquiry and interpreting a complex contract require different capabilities, output lengths, and costs. There’s no need to assume that every step must use an equally expensive general-purpose reasoning model.
My view is that businesses will increasingly prioritize dividing AI work by task. Simple decisions can go to suitable specialized models, while complex problems can be escalated for further handling. What really matters is comparing the cost and error rate across the entire workflow—not just how cheap a single model appears to be.
This also explains why usage volume doesn’t directly translate into revenue: as the number of calls grows, the mix of tasks and the price per call may change. If errors lead to repeated work, those apparent savings can disappear.
When looking at AI-related assets such as TAO, FET, and RENDER, it’s also worth asking what paid demand each one actually has. Microsoft’s product launch is not evidence that these tokens are generating revenue.
I’m more interested in where businesses are spending their budgets—and whether that step is genuinely worth paying for.
$TAO $FET $RENDER #AI
Tap my profile picture to view my live trades