The companies that moved fastest to adopt AI tools in 2025 aren't looking as smart as they did a year ago.

Uber burned through its entire 2026 AI coding budget in four months. Microsoft canceled most of its AI tool licenses after six months of unchecked spend. Nvidia's VP of research said compute costs had become "far beyond the costs of employees."

This isn't an AI failure. It's a procurement failure. And the fix is obvious — once you understand what's actually driving the bills.

The token consumption paradox

Here's the dynamic that's catching companies off guard: the cost per AI token keeps falling. So finance teams assume their AI bills will fall too. They don't.

Goldman Sachs forecasts a 24-fold increase in token consumption by 2030 as agentic AI becomes standard. Agentic AI — the kind that takes multi-step actions, loops through reasoning, calls tools, and checks its own work — consumes exponentially more tokens per task than simple prompt-response interactions.

Cheaper tokens × more tokens = bills that grow faster than budgets.

The math only works in your favor if consumption stays flat. It doesn't. Every new use case, every agent you deploy, every task you hand to an AI tool multiplies token usage — and the general-purpose platforms are designed to encourage exactly this kind of expansion.

Why general AI tools fail at scale

Off-the-shelf AI tools are built to handle anything, which means they handle nothing efficiently. An agentic tool completing a task your business runs 200 times a day will burn tokens on context-loading, intermediate reasoning, and self-verification — even when the underlying task is a three-step process that's identical every time.

The incentive structures make it worse. Platforms reward usage. Internal champions get credit for AI adoption metrics. Nobody's measuring whether the output per dollar is actually improving. By the time the overspend surfaces in a quarterly review, months of budget have already been consumed.

This is exactly what happened at Uber and Microsoft. The tools worked — technically. They just cost more to run than the human labor they were meant to replace.

What purpose-built automation does differently

Custom automation doesn't have a token problem because it doesn't use tokens.

A workflow built for one specific task — say, extracting data from inbound invoices and posting it to your accounting software — runs the same deterministic logic every time. It doesn't reason. It doesn't loop. It executes a defined sequence at near-zero marginal cost, regardless of volume.

That's not a limitation. That's the point. The workflow does exactly what you need, as many times as needed, without a meter running.

The right question to ask before adopting any AI tool

Not "how do we adopt more AI?" — but: what specific task costs us the most time or money right now, and can we automate exactly that?

If the answer is a narrow, well-defined workflow that repeats on a schedule or trigger — that's a custom automation candidate. It doesn't need a general AI platform. It needs a purpose-built workflow that runs predictably and cheaply in the background.

If the answer is something genuinely open-ended — creative work, research, complex reasoning — then a general AI tool may be the right fit. But that's a much smaller slice of most businesses' daily operations than vendors would have you believe.

The businesses getting durable ROI from AI right now aren't the ones running everything through a general-purpose agent. They're the ones who identified their most expensive manual tasks and replaced them with narrow, reliable automation that runs whether or not anyone's watching.