AI Costs Out of Control: How to Plan Your Token Budget

The moment the AI bill landed on the table

For years, the tech industry has been sprinting. The dominant logic was tokenmaxxing: maximize the use of language models, experiment, scale. But as a recent analysis by TechCrunch reports, the industry is now facing out-of-control AI infrastructure costs. The sentence that sums up the change says it all: «The whole conversation shifted from tokenmaxxing and ‘go fast’ to ‘we need guardrails, how do we control this?’»

So 2026 marks a real turning point. It's no longer about exploring what artificial intelligence can do. It's about understanding what it really costs to integrate it into business processes. This is true for large corporations, and even more so for Italian SMEs.

The numbers that matter: tokens, costs and real usage

The concept of a token is central to this analysis. A token corresponds roughly to three or four characters of text. Every call to a language model, whether GPT-4o, Claude 3.5 or Gemini, consumes tokens in both input and output. As a result, cost grows in proportion to the complexity and frequency of requests.

According to aggregated industry data, a mid-sized company that integrates AI into three or four workflows can end up spending between €2,000 and €15,000 a month on API calls alone. What's more, these costs tend to grow over time as use cases multiply. Without a monitoring system, spending becomes opaque and hard to justify in a budget review.

Research from McKinsey confirms that managing AI costs has become one of CTOs' top concerns in 2026. The difficulty is not just technical, it's organizational. Who approves the spending? Who monitors usage? Who decides when a model is too expensive for the value it generates?

The strategic view: why SMEs are more exposed than large companies

Large enterprises have dedicated MLOps teams. They have engineers who optimize prompts, shorten context length and pick the most efficient model for each task. Italian SMEs, by contrast, often adopt AI solutions in a fragmented way, relying on SaaS tools that hide the underlying costs.

This creates a specific risk. An SME that uses an AI writing tool, a customer service chatbot and a data analysis system may not realize it's paying for tokens three times over. In other words, fragmented tools mean fragmented costs, which are hard to add up and even harder to optimize.

On top of that, the AI model market is evolving fast. Newer, more capable models tend to cost more per token. Automatically upgrading integrations to the latest available version can therefore mean significant, often unplanned cost increases.

At SHM Studio, we see this pattern more and more often. SMEs that come to our AI services often have no clear map of their consumption. The first step of our work is always to build that visibility.

Anatomy of the problem: where costs hide

To know where to step in, it helps to map the main sources of token consumption in a typical company. There are three main spending categories.

• Inefficient prompt engineering: prompts that are too long, redundant context, instructions repeated on every call. Every extra character has a cost.
• A model that's oversized for the task: using GPT-4o to classify simple emails is like using a scalpel to slice bread. Smaller, specialized models cost less and often perform better on specific tasks.
• No caching: many implementations repeat identical or nearly identical calls without using cache mechanisms. This multiplies costs with no added benefit.

Likewise, the lack of spending limits set at the API account level is a real operational risk. Gartner has identified AI cost governance as one of the top ten technology priorities for 2026-2027, as you can read in its technology forecasts.

Operational implications for Italian SMEs

Turning this analysis into action takes a structured approach. Below are the priority areas for an SME that wants to bring its AI spending under control.

First, mapping. You need to take stock of every touchpoint with AI APIs: SaaS tools, custom integrations, automations. Without that visibility, any optimization is blind.

Next, segmentation by value. Not all AI tasks have the same strategic value. It helps to rank use cases by the return they generate and the cost they incur. That lets you allocate budget rationally.

Then, choosing the right model. Today there are open-source and mid-tier commercial models that deliver excellent results on specific tasks at a fraction of the cost of flagship models. Choosing a model is an economic decision as much as a technical one.

Finally, continuous monitoring. AI costs are not static. They grow with adoption, change with model updates and vary with usage volumes. A monitoring dashboard, even a simple one, is an essential tool.

Our digital marketing and SEO work increasingly includes AI components. Our clients are exploring these integrations too. That's why we've developed an approach that balances innovation with cost control.

What nobody tells you: the hidden cost of iterative prompting

There's one aspect that rarely comes up in discussions about AI costs: iterative prompting, the practice of gradually refining a prompt through dozens of test calls. In a team of three or four people working on an AI integration, this process can burn through hundreds of thousands of tokens before the system even goes into production.

Even so, few companies track these development costs separately from operating costs. The result is a systematic underestimation of the total cost of ownership of an AI integration. A company's real AI budget is often double what it perceives.

This doesn't mean AI isn't worth the investment. On the contrary, it means ROI needs to be calculated more rigorously. Our web services and our SEO copywriting show every day that AI, used well, creates measurable value. However,