AI sticker shock: why companies are spending millions on AI without seeing results
Over the past two years, the message has been clear: adopt AI or fall behind. Companies listened. They bought licenses, hired consultants, and rolled out copilots everywhere. And now, many are opening this month’s invoice and discovering something uncomfortable: AI is far more expensive than they thought, and the results aren’t coming.
Axios has dubbed it "AI sticker shock". And this isn’t a problem limited to small businesses with bad advice. It’s hitting the big leagues hard.
What’s happening right now
The headlines of recent weeks paint a clear picture. Microsoft canceled most of its Claude Code licenses, partly because of cost, according to The Verge, and Uber’s COO said AI costs are "increasingly difficult to justify."
Then there’s the case that went viral in tech circles: an AI consultant told Axios that one of his clients recently spent half a million dollars in a single month after failing to set usage limits on Claude licenses for employees.
Half a million. In one month. Because no cap was set.
It’s not an isolated case. One of the most striking examples of this spending surge came from Uber. The company’s CTO has already burned through the entire AI budget planned for 2026 because of token costs, according to a report from The Information.
And the statistic that best sums up the moment comes from Nvidia, a company that literally sells the shovels in this gold rush: Bryan Catanzaro, vice president of applied deep learning at Nvidia, put it bluntly: for his team, compute costs far exceed personnel costs.
If AI costs more than the people running it at Nvidia, imagine what’s happening at a regular company.
Why the bill spirals so fast
The problem isn’t that AI is expensive in itself. The problem is how it’s being used. Let’s break it down.
1. The all-you-can-eat myth
A lot of people came into AI thinking a monthly subscription was like paying for Netflix: use as much as you want and that’s it. That’s not how it works.
A CTO told Axios that his employees were using AI models to check the weather. That gets expensive very quickly: enterprise AI plans are not really "all you can eat," and even simple chatbot queries can rack up high token costs.
Every request consumes tokens. Every token costs money. Multiply that by thousands of employees making trivial requests all day long, and the bill shoots through the roof.
2. Automating the wrong things
This is probably the most expensive mistake of all, and it has nothing to do with technology.
"Most people default to automating tasks they dislike, instead of the tasks that are most valuable to the company," said Sophia Velastegui, CEO of Velastegui Ventures and former chief AI officer at Microsoft. Instead, they should focus on using AI to drive revenue.
That’s a massive distinction. There’s a huge difference between "I hate writing meeting notes, let AI do it" and "we lose 200 leads a month because we take 4 hours to respond, let AI qualify them and reply instantly."
The first saves you hassle. The second makes you money. Only the second justifies the cost.
3. The throw-licenses-at-the-wall strategy
Leadership doesn’t always help: throwing AI licenses at the wall to see what sticks (what Velastegui calls the "let a thousand flowers bloom" approach) isn’t producing tangible returns.
It’s the corporate equivalent of buying every tool in a hardware store without knowing what you’re building. You hand out Copilot, ChatGPT Enterprise, Claude, Cursor, Gemini... and by quarter-end you’ve got a six-figure bill and no one can say which specific problem got solved.
4. The human bottleneck
Humans are the bottleneck to more efficient adoption, because we’re still catching up to AI.
AI can be incredibly powerful, but if no one on the team knows how to integrate it into real processes, it just sits there as another tool on the desktop. It gets used for isolated tasks, not to transform workflows.
5. The fear of opening up data
When companies hesitate to give AI agents unrestricted access to proprietary data, those agents become less effective, according to Josh Pantony, CEO of Boosted.ai.
That’s the paradox of the moment: the more you lock data down for security, the less useful AI becomes. The more you open it up, the more nervous you feel. Without a clear strategy for who can access what and why, you end up stuck in no man’s land.
"Tokenmaxxing" and the correction ahead
The market is already reacting. The industry is going through a "healthy correction" away from excessive AI use — or "tokenmaxxing," the drive to burn as many AI tokens as possible — according to Ali Ansari, CEO of Micro1.
And here’s the stat that should worry any executive most: even though the market treats these tools as if they work equally well across the entire company, Ansari says that "the reality of AI right now is that it only really works for coding." That disconnect can drive IT bills up without generating a high return on investment in agents.
Important caveat: that statement isn’t entirely true — there are plenty of use cases outside coding where AI, when applied well, works brilliantly (customer support, email management, lead qualification, voice agents). But the key point is this: it works well when it’s designed properly for a specific use case. Not when it’s handed out like candy.
The real consequences
This isn’t some abstract spreadsheet problem. It affects people.
Companies are citing AI’s ability to automate jobs as a reason for layoffs, even though Anuj Kapur, CEO of CloudBees, told Axios that workforce cuts may simply be "the only lever they can pull" to offset their AI bills.
Read that again. Some companies are laying people off not because AI made their jobs redundant, but because they need to pay the AI bill somehow. That’s the exact opposite of what was promised.
Consumer sentiment toward AI is also dropping fast, and employees are rebelling against the technology’s use at work.
How to do it right: the opposite playbook
Okay, enough diagnosis. What should you do? If you’re reading this from the chair of a founder, COO, or tech lead, here’s the short version of what works and what doesn’t.
| What does NOT work | What DOES work |
|---|---|
| Buying licenses for everyone and hoping for the best | Identifying 1-2 specific processes with measurable cost |
| Automating annoying tasks | Automating tasks that generate revenue or save measurable time |
| Paying premium model pricing for everything | Using the right model for each task (most don’t need GPT-5) |
| Generic "for companies like yours" solutions | Solutions designed for your exact processes |
| Getting locked into a closed vendor | Working with open-source tools and APIs you can switch out |
| Implementing and forgetting | Monitoring, measuring, and continuously adjusting |
Start with the problem, not the tool
It sounds obvious. It isn’t. Most AI projects start like this: "we want to implement AI." That sentence is already the first mistake.
The right question is: what specific process is costing us time or money, and can we measure it?
Real examples:
- "We spend 3 hours a day handling support emails"
- "We lose 40% of leads because we don’t respond fast enough"
- "Building a sales proposal takes us 2 days"
- "We handle 200 calls a month that are always the same 5 questions"
Each of these problems has a concrete monthly cost. And each has an AI solution that can be designed to solve it specifically, with ROI that can be calculated from day one.
Set limits from the start
The case of the client who spent half a million dollars comes down to a very simple reason: nobody set a cap. In any well-designed solution, there are hard limits: how many calls can be made, which model is used for what, when to escalate to a human, when to stop.
Measure ROI in real time
If you can’t answer, "How much has this system saved us this month?" then you don’t have an AI system. You have an expensive experiment.
The Studio SmartWork approach
This whole article basically explains why Studio SmartWork exists.
We don’t sell licenses. We don’t sell all-in-one software. We don’t tell you to "try it and see."
What we do is:
- We start with the specific problem. A quick audit of the workflow where you’re losing the most time or money. Real numbers: how many hours, how many leads, how many euros per month.
- We design a custom solution. Using proven open-source tools (n8n, AI APIs), integrated with your current stack. No vendor lock-in.
- We run it. We don’t hand you a manual and disappear. We monitor, maintain, and improve the system as your business grows.
- In 4-8 days. Not 6 months. Not a quarter. In less than a workweek.
The results are measurable from the first month: inboxes that go from 3 hours to 15 minutes a day, lead response times that drop from hours to under 60 seconds, sales proposals that go from 1-2 days to 10 minutes.
Not because we use AI for the sake of using AI. Because every solution is designed for a specific process, with clear limits and clear metrics.
The bottom line
The "AI sticker shock" story isn’t about AI not working. It’s about buying AI without a strategy not working. Those are very different things.
Applied well, AI remains one of the most powerful levers a business can use. But "applied well" means: to a specific problem, with measurable ROI, clear boundaries, and someone actually operating it. Not "I hand Copilot to 200 employees and see what happens."
What we’re seeing is whether companies become more disciplined with AI use. Or whether they overcorrect and shut themselves off.
The smart answer is neither. It’s to be surgical: use AI where the economics make sense, don’t use it where they don’t, and measure everything. That’s the difference between the companies that will be sharing success stories in 2027 and the ones that will be explaining to the board why their AI bill exceeded payroll.