AI Pulse · Business Models & Winners

Microsoft just shipped a security model that handles about 90% of routine cybersecurity tasks on its own and only escalates the tricky 10% to a more expensive frontier model. The result is roughly half the cost of running everything through the premium blend it replaces, according to the company’s announcement.

This matters because it’s a working example of the cost architecture most businesses will need if they want AI to stay profitable at scale. You train or tune a smaller, cheaper model to handle the repetitive stuff: triaging alerts, classifying threats, generating standard responses. When it hits something genuinely ambiguous or high stakes, it kicks the decision up to a bigger, smarter, more expensive model. The bulk of the work stays cheap. The margin stays intact.

If you’re running AI on a fixed-fee retainer or trying to keep usage predictable, this is the pattern that makes the math work. A command centre built this way, like what we assemble in Omni, routes tasks by complexity so you’re not burning frontier-model tokens on questions a fine-tuned assistant could answer in its sleep. You get the performance where it counts and protect your margin everywhere else.

The broader lesson is that stacking models by capability, not just picking one and running everything through it, is how you avoid cost blowouts when volume climbs. Microsoft is doing it for security workflows. The same logic applies to customer support, document review, lead qualification, or any process where most inputs are routine and a few need real judgment. Build the stack to match the distribution of difficulty, and you stop paying luxury prices for economy work.