Rising AI Costs: How Italian SMEs Are Managing Tokens

The context: from token euphoria to the reality of costs

In 2023-2024, the adoption of generative AI followed a logic of maximum acceleration. Companies experimented, integrated and scaled. Language model providers also cut prices to win market share. The result was an environment where cost per token seemed like a minor issue.

But things have changed radically. In 2025, many organizations discovered that their AI bills had grown exponentially. According to a recent analysis by TechCrunch, the industry hit a real breaking point. The phrase going around technical teams says it all: “The whole conversation shifted from tokenmaxxing and ‘go fast’ to ‘we need guardrails, how do we control this?'”.

So 2026 is the year AI's economic sustainability became a strategic priority. And not only for large enterprises, but also for Italian SMEs that have built LLM-based tools into their operations.

The numbers that matter: what AI really costs an SME

Quantifying the cost of AI for a small or mid-sized business isn't simple. Spending is spread across several items: platform subscriptions, API costs for tokens consumed, development hours for integration, and prompt maintenance. As a result, the actual budget is often higher than planned.

Some market estimates give a useful picture. According to Gartner research, by 2027 more than 40% of organizations adopting generative AI will exceed their initial budgets. Proactive cost management is therefore becoming a critical skill.

For an Italian SME with 20-100 employees, monthly AI costs can range from a few hundred to several thousand euros. It depends on request volume, prompt complexity and the choice of model. Latest-generation models like GPT-4o or Claude 3.5 in particular have significantly higher per-token costs than earlier ones.

There are also hidden costs. Take context tokens: every time you send a long conversation to a model, you pay for all the previous tokens too. This can multiply your spending in ways you wouldn't expect.

Strategic view: why the problem is structural

The AI cost problem isn't temporary. Nor will it solve itself by waiting for prices to fall. Competition between providers has already squeezed margins. Further price cuts will be gradual and won't offset the growth in usage.

So the issue is structural. Companies that build AI into their workflows tend to consume more over time. This is known as usage creep: every new use case adds tokens, every automation generates new requests. Without clear governance, spending grows organically and becomes hard to control.

An analysis by Harvard Business Review points out that the most effective organizations in AI adoption aren't the ones that spend the most. Instead, they define priority use cases up front and measure the return on each application. Strategic discipline matters more than budget.

For Italian SMEs, this principle is even more relevant. Margins are often tighter. In-house technical resources to monitor and optimize AI spending are also limited. In short, you need a methodical approach that not every company has developed yet.

The work still ahead: governance and operational guardrails

The industry's response to the cost problem works on several levels. First, many companies are implementing granular monitoring systems. Every API request is tracked, classified by use case and assigned to a cost center. That makes it possible to spot where money is being wasted.

Next, they work on prompts. Prompt engineering isn't just about output quality. It's also a lever for cost optimization. More precise, concise prompts produce better answers while using fewer tokens. Investing in this skill pays off twice.

Many organizations are also reassessing their choice of model. The most powerful model isn't always the right one. For simple tasks like classification, data extraction or short text generation, lighter and cheaper models deliver equivalent results. A model routing strategy, which assigns each task to the most suitable model, can cut costs by 30-50% without affecting quality.

At SHM Studio, we work with our SME clients on exactly this kind of optimization. The goal isn't to use less AI. It's to maximize the value generated for every euro spent.

Operational implications for B2B retail and professional services

The sectors most exposed to AI cost problems in Italy are B2B retail and professional services. Both have adopted AI tools to automate communications, generate content and support customer service. But they've often done so without structured governance.

For B2B retail, the most common use cases include generating product descriptions, personalizing offers and supporting negotiations. Each of these processes consumes tokens continuously. Companies with large catalogs and intense sales cycles can therefore rack up significant costs.

For professional services, such as law firms, accountants and communication agencies, AI is used to draft documents, summarize information and generate reports. Here the main risk is context window overload: long documents sent repeatedly to the model multiply costs exponentially.

The operational implications are clear. You need an audit of active use cases. You need a map of real costs. And you need a prioritization strategy that separates high-ROI uses from marginal ones. Companies that start this process today will have a significant competitive advantage over the next 18-24 months.

Tools and approaches to optimize AI spending in 2026

There are already tools on the market dedicated to monitoring and optimizing AI costs. Among the most widely used are LLM observability solutions such as LangSmith, Helicone and Portkey. These tools let you track every API call, measure latency and calculate the cost per output generated.

However, adopting these tools requires technical skills that many SMEs don't have in-house. That's where an experienced digital partner makes the difference. An agency like SHM Studio can support your company with the audit, tool selection and the definition of sustainable usage policies.

Beyond that, some architectural approaches cut costs at the structural level. For example, caching responses to recurring queries means you don't have to call the model every time. Similarly, fine-tuning smaller models on company-specific data can deliver results comparable to large models at a fraction of the cost.

If you run digital marketing or SEO activities, optimizing AI costs also means rethinking your content production workflows. In particular, defining standardized prompt templates and limiting the length of the context sent to the model are quick, high-impact fixes.

The SHM Studio perspective: ROI first

The debate on AI costs risks splitting into two extreme positions. On one side, those who argue that you need to cut AI use to contain spending. On the other, those who believe any cost is justified by innovation. Both positions are wrong.

The right perspective is ROI. Every AI application should be judged by the value it generates. So the question isn't “how much are we spending on AI?” but “how much is every euro spent on AI worth?”

To answer that question, SMEs need to build specific metrics. For content generation, for example: cost per article produced with AI vs. cost with the traditional method, plus a comparison of quality. For automated customer service: cost per resolved ticket and first-contact resolution rate. For ad campaign management: cost per qualified lead generated with AI support.

Finally, it's important to consider opportunity costs. A company that doesn't optimize its AI spending today risks losing ground to competitors that do. So optimization isn't optional. It's a strategic necessity.

If you want to dig deeper into these topics, you can explore the resources on the SHM Studio blog or contact us directly through the contact page for a tailored consultation.