Microsoft reveals the true cost of AI: why adoption is plateauing despite billions invested

Microsoft reveals the true cost of AI: why adoption is plateauing despite billions invested

🔎 AI was supposed to reduce costs — it's making them explode

May 2026. Microsoft publishes financial reports that sound like a warning. AI agents and token consumption cost, in some cases, more than the salaries of the humans they are meant to replace.

This isn't being said by some anonymous doomsayer. It's the company that invested $13 billion in OpenAI and is building its entire strategic empire around generative AI.

Meanwhile, in Europe, companies continue to massively ignore these tools. And the Stanford AI Index 2026 report confirms a growing gap between expert forecasts and reality on the ground.

The promise was simple: AI would lower operational costs. The reality is more complex, more expensive, and much less linear.

The Essentials

• Microsoft reports (May 2026) reveal that AI token and agent costs often exceed the cost of equivalent human employees.
• Uber consumed its entire 2026 budget for AI coding tools in just four months.
• Microsoft cancels most of its Claude Code licenses due to rising costs, while negotiating to supply its Maia chips to Anthropic.
• Salesforce plans to spend $300 million on Anthropic tokens in 2026, and is calling for a "routing layer" to reduce the bill.
• Eurostat 2025 shows that European companies are still holding back: lack of skills, data confidentiality, and legal uncertainty.
• The Stanford HAI AI Index 2026 confirms that agent success rates have tripled (from 12% to 77%), but perspectives diverge between experts and the general public.
• If your digital strategy relies on AI without mastering its real costs, your website without AI may already be obsolete in 2026 — but the reverse is also true: poorly calibrated AI can ruin you.

Recommended Tools

What Microsoft's reports really reveal

AI costs more than humans in several concrete scenarios. This is the stark conclusion of Microsoft's internal documents analyzed by Fortune on May 22, 2026.

The Fortune article details how Microsoft revealed, through its financial reports, that the use of AI agents and token consumption reached cost levels higher than paying human employees for equivalent tasks.

The problem is not theoretical. Uber burned through its entire 2026 budget dedicated to AI coding tools in just four months. Four months for an annual budget.

Microsoft itself has taken drastic measures: the company is canceling most of its Claude Code licenses, Anthropic's product for AI-assisted code. The reason given is direct — the continuous rise in token costs makes the operation unprofitable at scale.

It is ironic. Microsoft, which invests billions in compute via Azure and its OpenAI partnerships, finds itself forced to ration its own AI consumption.

The compute paradox

Microsoft does not lack computing power. The company is developing its own Maia chips and negotiating to supply them to Anthropic as an alternative to NVIDIA GPUs.

But computing power does not solve the economic problem. Every request, every autonomous agent, every reasoning loop consumes tokens. And tokens have a price that, at enterprise scale, becomes astronomical.

Livemint reports in late May 2026 that Microsoft and Uber both illustrate the same phenomenon: AI was supposed to reduce costs, but it ends up costing more than the human employees it replaces or assists.

The Salesforce case: 300 million dollars in tokens

If you thought the problem was limited to Microsoft and Uber, look at Salesforce.

Marc Benioff, CEO of Salesforce, has publicly announced that his company plans to spend 300 million dollars on Anthropic tokens for the year 2026. This figure, reported by The Next Web, includes the integration of AI into Slack and internal coding tools.

Three hundred million dollars. In tokens alone. Not in infrastructure, not in engineers' salaries, not in training. Solely in requests sent to Anthropic's models.

The "routing layer" as a rescue plan

Benioff isn't just announcing this figure — he is calling for an architectural solution. He advocates for a "routing layer," a middleware that would automatically dispatch requests between frontier models (like Claude Opus 4.7) and smaller, cheaper models.

The idea is appealing on paper: why send a simple question to a model that costs 10 times more than a model sufficient for 90% of daily tasks?

But this routing layer does not yet exist in a standardized way. And building it requires skills that most companies do not have — which brings us back to the European problem.

Europe: why adoption remains a desert

While US giants battle with AI costs, Europe faces a different problem: it can't even adopt these tools.

Euronews, in a May 25, 2026 article drawing on Eurostat 2025 data, identifies three major barriers to AI adoption by European companies.

The first is the lack of technical skills. European companies, particularly SMEs, do not have the ML engineers or data scientists needed to deploy and maintain AI solutions.

The second barrier concerns data confidentiality. European regulations, notably the GDPR, make companies extremely cautious about sending sensitive data to external APIs. This is a structural barrier that the European AI Act attempts to clarify, but with effects that remain to be measured.

The third barrier is legal uncertainty. The regulatory framework is evolving, companies are waiting, and waiting is costly in terms of competitiveness.

The gap between countries

The Economist Impact study, citing Eurostat 2025 data and the Federal Reserve Bank of St. Louis (2026), shows that differences in adoption between countries are considerable.

Even with the drop in token prices observed between 2024 and 2026, structural barriers persist. Access to compute, team training, basic digital maturity — all of this varies enormously from one European country to another.

It is not a problem of willingness. It is a problem of ecosystem.

Stanford AI Index 2026: the gap between hype and reality

The Stanford HAI publishes its AI Index Report every year, which has become the global benchmark for measuring the state of AI. The 2026 edition, analyzed by ArtificialStudio, confirms a paradoxical trend.

AI agent success rates have tripled, going from 12% to 77% on real-world tasks. This is a considerable technical leap. Agentic models like GPT-5.5 (agentic score of 98.2) or Gemini 3 Pro Deep Think (95.4) are objectively more performant than they were a year ago.

But — and this is a massive but — perspectives diverge sharply between experts and the general public regarding the short-term economic impact. The 5 figures from the Stanford AI Index 2026 clearly show it: the technology is advancing, the economy is struggling to keep up.

Experts vs. general public: two parallel visions

Experts, who are close to the models and the infrastructure, see the progression curve and project upcoming profitability. The general public and business leaders, on the other hand, see rising bills and productivity gains that are difficult to quantify.

This gap is not new in the history of technology. But with AI, the scale of the investments makes the risk financially significant for companies that jump in without a safety net.

The cost hierarchy: which models for which use cases?

Not all models are created equal, and above all, they don't all cost the same price. Understanding this hierarchy is essential to avoid the financial pitfalls that Microsoft and Uber encountered.

Generalist models: when the top tier isn't necessary

The June 2025 ranking of generalist LLMs shows a clear hierarchy. Gemini 3.1 Pro (Google) dominates with a score of 92, followed by GPT-5.5 and GPT-5.4 Pro (OpenAI) at 91.

But for an email classification task, meeting summarization, or content moderation, a Claude Sonnet 4.6 (score 83) or a GLM-5.1 (83) is more than enough. The extra cost of a frontier model brings nothing measurable.

Agentic models: where costs really explode

It is in the agentic domain that costs become dizzying. Autonomous agents multiply API calls — each reasoning step, each tool called, each verification consumes tokens.

GPT-5.5 dominates the agentic ranking with 98.2. But each complex task entrusted to this agent can generate dozens, or even hundreds, of nested calls. At enterprise scale, the bill explodes.

This is exactly what happened at Uber with its AI coding tools. A coding agent doesn't call the model just once — it iterates, tests, corrects, starts over. The cost per final task can be multiplied by 10 or 20 compared to a classic "chat" usage.

Comparison of indirect costs by task category

AI Agents vs chatbots: the price of autonomy

The distinction between a simple chatbot and an AI agent is not just technical. It is fundamentally economic.

An AI avatar is not a chatbot, and this difference translates directly into costs. A chatbot answers a question with a single API call. An agent plans, executes, iterates, and each step costs.

When Microsoft talks about AI agents that cost more than humans, it is precisely this mechanism being discussed. Autonomy has a price — and in 2026, this price is not yet offset by productivity gains for the majority of companies.

Why productivity gains don't offset the costs

Economic theory is simple: if AI does in 10 minutes what a human does in 2 hours, even a high hourly cost for AI should pay for itself.

In practice, it's different for several reasons.

First, productivity gains are uneven. Stanford HAI notes that experts and the general public have very different perspectives on the economic impact. A senior developer might save 30% of their time. An administrative employee might lose time verifying and correcting AI outputs.

Second, invisible costs accumulate. Training teams, integrating into existing workflows, maintaining AI pipelines, monitoring output quality — all of this costs time and money that Microsoft reports don't always factor into the "cost of AI."

Finally, the rebound effect is real. When something becomes easier and cheaper (seemingly), we do more of it. Teams generate more content, more analyses, more code — and token consumption follows an exponential rather than a linear curve.

The marginal cost trap

The marginal cost of an additional AI request is close to zero — this is the classic argument of API vendors. But "close to zero" multiplied by millions of requests yields a number very far from zero.

Salesforce with its $300 million is the perfect illustration. Each individual token costs almost nothing. The cumulative bill is colossal.

What this means for the cost of a website in 2026

AI isn't just affecting large enterprises. It is also transforming the economics of web creation.

Integrating an AI chatbot, a search assistant, or generative features into a website carries a recurring cost in tokens that didn't exist two years ago. The true price of a website in 2026 is no longer just the initial development — it's also the monthly AI bill that comes with that site.

An SMB that integrates a Claude Sonnet 4.6 agent for customer support could see its monthly bill double or triple compared to a classic static website. Without an efficient routing layer — the kind Benioff is calling for — these costs are difficult to control.

❌ Common mistakes

Mistake 1: Deploying a frontier model for simple tasks

Using GPT-5.5 (agentic score 98.2) to answer basic FAQs is like using a sledgehammer to crack a nut. The cost per request is disproportionate to the value generated.

Solution: Establish a task/model matrix. Route simple tasks to Claude Sonnet 4.6 or GLM-5.1, and reserve frontier models for cases that objectively justify them.

Mistake 2: Ignoring the cost of agentic loops

An agent that iterates 15 times to solve a problem doesn't cost 15 times more — it potentially costs 50 times more because of the context tokens accumulated at each iteration.

Solution: Set iteration limits per task. Monitor the real cost per completed task, not per individual request.

Mistake 3: Underestimating European compliance costs

European companies that dive into AI without integrating GDPR and AI Act constraints from the design phase end up with costly refactoring or abandoned projects.

Solution: Integrate compliance into the initial specifications, not as an add-on. The cost of proactive compliance is a fraction of the cost of reactive compliance.

Mistake 4: Believing that falling token prices solve everything

Economist Impact notes that despite falling token prices, structural barriers persist in Europe. Price is just one factor among others.

Solution: Treat token cost as a lever, not a solution. Skills, data governance, and technical architecture remain prerequisites.

❓ Frequently asked questions

Is AI really more expensive than humans?

In specific scenarios documented by Microsoft (May 2026), yes. AI coding agents consumed Uber's annual budget in 4 months. But this depends heavily on the type of task, the model used, and the efficiency of the routing between models.

Why are European companies adopting AI so little?

According to Eurostat 2025, three main barriers dominate: the lack of technical skills, data privacy concerns (GDPR), and the legal uncertainty linked to the evolving regulatory framework.

How much will Salesforce spend on AI in 2026?

300 million dollars on Anthropic tokens alone, according to Marc Benioff's announcement reported by The Next Web. This figure does not include infrastructure, salaries, or other AI providers.

Is Benioff's routing layer a real solution?

It is a logical architectural approach — dispatching requests to the most suitable and cheapest model. But it requires advanced ML ops skills that most companies do not yet have.

Have AI agents become more reliable?

Yes. The Stanford AI Index 2026 shows that success rates have gone from 12% to 77% on real-world tasks. But a 77% reliability rate still means 23% failure or manual correction — a significant hidden cost.

Which model to choose to get started without breaking the bank?

DeepSeek V4 Pro (High) with a score of 84 as a generalist and a cost per token significantly lower than American models, or Claude Sonnet 4.6 (83) for a good cost/performance balance. To be reserved for tasks that justify it.

✅ Conclusion

AI in 2026 is a technology that has tripled its performance in two years — but has not yet solved its economic equation. The Microsoft reports, Salesforce's $300 million, and European hesitations all paint the same picture: the technology is there, the business case isn't always.

If you are building or evolving your digital presence, choose an infrastructure that integrates AI at a controlled cost — and above all, don't confuse technical capability with profitability.