We’ve spent two years hearing that AI is going to make everything cheaper. That it’s going to replace workers. That it will be the biggest productivity revolution in history.
And then, this week, Fortune published a story that feels like it came from a joke: using AI in a company’s day-to-day operations can cost more than keeping human employees on the payroll.
This isn’t coming from a skeptic on the outside. It’s coming from Microsoft, Uber, and even a vice president at Nvidia — the company making the most money from all of this.
Let’s break down what’s happening, why it’s happening, and — more importantly — what it means for your business if you’re thinking about automating with AI.
What has happened in recent weeks
Here’s the short version:
- According to The Verge, Microsoft has started canceling most of its direct Claude Code licenses and is moving its engineers to GitHub Copilot CLI. This comes just six months after the company opened access to Claude Code to thousands of developers, project managers, designers, and other employees.
- Uber CTO Praveen Neppalli Naga admitted that its reliance on Claude Code had already exhausted the annual AI budget just four months into 2026.
- "For my team, compute cost is very much over and above employee cost," Bryan Catanzaro, Nvidia’s vice president of applied deep learning, told Axios.
When the executive from the company that sells the shovels in this gold rush tells you the shovels cost more than the miners, it’s worth stopping and listening.
The token paradox: cheaper per unit, more expensive overall
This is where things get interesting. The usual narrative says AI keeps getting cheaper. And technically, that’s true:
A recent Gartner report found that by 2030, inference in a one-trillion-parameter LLM will cost AI companies nearly 90% less than it did in 2025.
So where’s the problem? In usage.
Gartner predicted that cheaper tokens will not translate into cheaper enterprise AI because agentic models require far more tokens per task than standard models, rising consumption can outpace falling unit costs, and AI providers will not fully pass lower costs on to consumers. In turn, inference costs are likely to rise.
Translated into business language:
| What we expected | What is happening |
|---|---|
| Cheaper tokens → lower bill | Cheaper tokens → we use far more → higher bill |
| One agent = one cheap digital worker | One “agentic” agent uses up to 1000x more tokens than a standard model |
| Obvious ROI from day one | Hard-to-prove ROI, budgets blown |
Goldman Sachs recently predicted that agentic AI could drive a 24x increase in token consumption by 2030, reaching a staggering 120 quadrillion tokens per month.
More efficiency does not mean lower cost. It means more usage. And in a pay-per-use model, more usage = a bigger bill.
The “tokenmaxxing” phenomenon: when using AI becomes an end in itself
There’s an almost comic side to this story. Big tech companies have pushed employees so hard to use AI that an absurd internal competition has emerged.
Uber and Microsoft are not the only companies pushing employees to use as much AI as possible. Just like at Uber, a Meta employee created a ranking, aptly named "Claudeonomics" after Anthropic’s AI model, to track which workers are using the most AI. Amazon is pressuring its employees to “tokenmaxx,” meaning to use as many AI tokens as possible.
The slang for this is “tokenmaxxing,” with some heavy users piling up monthly token bills of more than $150,000. “I probably spend more than my salary on Claude,” Max Linder, a software engineer in Stockholm, told The New York Times last month.
It’s the modern equivalent of buying tools just because they’re on sale, even if you don’t need them. And worse: this phenomenon has led many employees to use AI for almost anything just to hit internal targets. This was evident at Amazon, where some team members admitted to using the tool for unnecessary tasks and inflating internal usage scores.
Why should you care if you run an SME?
Here’s the point. If you manage a company with 10, 50, or 200 people, this is not a “big tech problem.” It’s an early warning sign you need to understand before diving into AI automation.
The underlying mistake Microsoft, Uber, and others are making is the same one many SMEs are about to make:
- Adopting AI as a goal, not as a means. Using AI because “we have to use AI,” without a clear business case.
- Choosing pay-per-use models without understanding the real cost. The price per token looks tiny. The monthly bill does not.
- Letting “agents” run without supervision or limits. A poorly designed agent can burn through thousands of euros before Friday.
- Getting locked into a single expensive provider. When the provider raises prices — and it will — you have no escape.
Token pricing follows the familiar trap: affordable at first to create dependence, then prices rise once companies are hooked.
How to do it right: AI that creates value without blowing up your budget
The Microsoft story is not “AI doesn’t work.” It’s “badly designed AI costs a fortune.” There’s a huge difference.
At Studio SmartWork, we’ve been building AI for businesses since 2021, and our approach is designed specifically to avoid these problems. Here are the principles we use — you can apply them yourself, or ask us to apply them for you:
1. Start with the problem, not the tool
Nobody should install AI because it’s trendy. Before touching a model, you need to understand:
- What process is wasting time or money?
- How much does it cost today in hours, errors, or lost opportunities?
- What would the real savings be if we automated it?
If you can’t answer those three questions, you’re not ready to invest in AI. You’re ready for a process audit.
2. Use the smallest model that solves the problem
You do not need GPT-5 or Claude Opus to classify emails or score leads. The obsession with always using the most powerful model is what is blowing up budgets at big tech companies.
Good AI design uses small, inexpensive models for 80% of tasks, and only brings in powerful models when complex reasoning is actually needed. That can cut your bill by 10x or even 50x.
3. Build deterministic workflows, not loose agents
An “autonomous agent” that decides what to do at each step sounds impressive. In practice, it’s what burned Uber’s annual budget in four months.
A well-designed workflow with defined steps, validations, and AI only where it adds real value is:
- More predictable
- Much cheaper
- Easier to debug
- Easier to maintain
That’s exactly what we do with n8n and purpose-built AI models: robust workflows that recover on their own when something fails, not agents burning tokens in circles.
4. Open source whenever possible
Microsoft is discovering just how expensive it is to depend on an external provider (Anthropic) when usage scales. The same thing can happen to any SME that builds its entire operation on a single proprietary API.
We work with open-source tools (mainly n8n) precisely for this reason: the client is never locked in. If the AI provider raises prices tomorrow, we can switch models without rewriting the whole system.
5. Measure ROI from day one
Companies now demand proof of ROI instead of demos. That should be your mindset from the start.
Every AI solution you implement should have clear metrics:
- Human hours saved per week
- Monthly cost of the solution (tokens + infrastructure + maintenance)
- Economic value generated (extra sales, errors avoided, faster response times)
If you can’t calculate ROI, it’s not an investment. It’s an expense.
The uncomfortable conclusion
The lesson from Microsoft is not that AI is useless. It’s that AI without design is incredibly expensive. And that letting agents run without strategy is like leaving the tap open and acting surprised when the bill arrives.
The good news: for a well-advised SME, AI is still one of the most powerful levers available. Eliminating 4 hours a day of manual lead follow-up, cutting an inbox from 3 hours to 15 minutes, or going from 3 sales proposals a week to 10 — all of that is still perfectly achievable, and at a very reasonable cost.
The difference between Microsoft burning millions on Claude Code and an SME automating with common sense is exactly this: design.
You don’t need the most powerful model. You need the right model, in the right place, with the right limits, measuring the right things.
If you want to understand which processes in your business can be automated with AI — and, above all, which ones are economically worth it and which ones are not — that is exactly what we do in the first conversation with you. No empty demos, no tokenmaxxing, no endless contracts. Just the problem, the solution, and the numbers.