The Signal #002 — Dakota’s read on the AI news that actually matters to people running a business.

There’s a story going around that sounds like bad news for AI. It’s actually one of the most useful things an operator can understand before spending a dollar on it.

Microsoft’s own internal reports, as covered by Fortune, show that running AI agents can cost more than paying the humans who used to do the same work. Not someday. Right now.

Here is what that means for your business, and where AI still wins anyway.

What happened

Last time in The Signal, we looked at why Microsoft pulled Claude Code from thousands of its own employees. The org-chart story was one layer. The cost reports underneath it are the more useful one.

After six months of encouraging staff to use AI coding tools heavily, Microsoft started reversing course, reportedly because of how expensive that usage got at scale. They are not alone. Uber’s CTO, Praveen Neppalli Naga, told The Information in April that the company burned through its entire 2026 AI coding tools budget in four months. An Nvidia executive, Bryan Catanzaro, put it plainly: “For my team, the cost of compute is far beyond the costs of the employees.”

The reason is how most AI tools are priced. You pay per token (the small chunks of text a model reads and writes). That sounds cheap per use. The problem is the bill scales with usage, and it doesn’t level off. Goldman Sachs projects token consumption could grow 24 times over by 2030. Gartner expects the price per unit to fall nearly 90% in that same window, and total enterprise AI bills to climb anyway, because everyone simply uses more.

That is the trap. Cheaper per use does not mean cheaper. It usually means more use.

Why it matters for operators

You are not Microsoft, and you are not running thousands of AI agents. That is exactly why this matters: at your scale, the cost question is simpler, and you can actually answer it.

The headline “AI costs more than employees” is true for some work and false for other work. The deciding factor isn’t AI versus humans in the abstract. It is the value of one run against the cost of one run, at your real volume.

AI wins on cost when a single use is short, the volume is spiky, and the thing it replaces is genuinely expensive or impossible. The clearest example is your phone. A missed after-hours call is a lost job worth hundreds or thousands of dollars. An AI receptionist only runs in short bursts, when a call actually comes in, and the alternative is a 24/7 answering service or leads that quietly vanish. The math is lopsided in AI’s favor.

AI loses on cost when a task runs constantly, every run is low value, and a person or a cheap tool already handles it fine. Pointing a top-tier model at huge volumes of routine work all day is exactly the pattern that blew up Microsoft’s and Uber’s budgets. At enough volume, you are paying premium compute to do a clerk’s job.

What most people get wrong

Most people assume AI is automatically the cheap option. “It’s a computer, computers are cheap.” So they either avoid AI entirely after reading a headline like this one, or they turn it loose on everything and get a surprise bill.

Both miss the point. The same tool can be a bargain for one task in your business and a money pit for another, on the same day. The skill isn’t “use AI” or “don’t use AI.” It is knowing which jobs clear the bar.

A simple gut check before you automate anything:

How often does this run, and how valuable is each run?
What does the failure, or the missed opportunity, actually cost me today?
Could a cheaper tool or someone already on payroll do it about as well?

If it is high value, spiky, and currently slipping through the cracks, AI usually wins. If it is high volume, low value, and already covered, AI usually loses, at least at today’s prices.

The short version

Microsoft and Uber found out the expensive way that AI billed per token gets pricier the more you lean on it. That is real, and it is worth respecting.

It does not mean AI is too expensive to use. It means you pick your spots. Put it where one run is cheap and one miss is costly, like your phone and your follow-ups. Keep it away from running flat out on work that was never worth much per task.

Cheaper per use is not the same as cheaper. Decide where AI earns its keep, and the headline stops being scary.

If you want help figuring out which jobs in your business actually pencil out for AI, that is what we think about at xovionlabs.com.