The AI spending slowdown: when costs catch up with enthusiasm

Some companies that deployed generative AI solutions at scale over the past two years are now actively limiting their use. Not for technical or ethical reasons, but for a more straightforward cause: the invoices don't fit the budget. According to a report from the Financial Times published on 19 June 2026, several technology leaders describe the situation with a phrase that has become recurring in internal conversations: "We created a monster".

The FT article documents how organisations across different sectors (finance, consulting, services) have begun imposing controls on which employees can use AI tools, how often, and for what tasks. Some have established monthly token quotas or withdrawn licences from teams whose return wasn't justifiable.

From rapid adoption to spending management

The pattern is recognizable for anyone who has followed the sector closely. Between 2023 and 2025, many companies deployed integrations with language models (via direct API, through intermediary platforms or tools like Claude Code) with a logic of "deploy first, measure later". The argument was that the cost of not adopting exceeded the cost of experimenting. That calculation is being revisited.

The issue isn't just the price per token, which in absolute terms has dropped considerably with successive generations of models. The issue is volume. When an organisation embeds AI in everyday workflows—writing, document analysis, customer service, code generation—consumption scales non-linearly. A fifty-person team using the model multiple times a day can generate a monthly bill that wasn't anticipated in any 2024 departmental budget.

Add to this the tendency to use more capable models for tasks that don't require them. Calling Claude Opus 4.8 with a 1M token context window to summarize an email is technically possible, but economically absurd. Optimizing which model to use for each task (something that in well-configured environments is managed with routing logic) remains an unresolved challenge for most companies.

What companies are doing to control spending

The measures described by the FT and those we've seen applied in integration projects go in several directions:

• Access restriction: limiting which profiles have access to which models, reserving the most expensive ones for justified use cases.
• Usage quotas: establishing monthly limits per user or team, with alerts when approaching the threshold.
• ROI audits by use case: reviewing which integrations generate measurable value and which are simply convenient but dispensable.
• Intelligent routing: using lighter models—Claude Haiku 4.5 or Sonnet 4.6—for low-complexity tasks, and reserving Opus for analysis that truly warrants it.
• Caching and context reuse: leveraging prompt caching features available in the API to reduce the number of tokens processed in repetitive conversations.

None of these measures is new or particularly sophisticated. What's striking is that many organisations didn't apply them from the start.

Why this matters beyond cost

The debate about AI spending has a dimension that goes beyond accounting. When a company restricts access to tools that its teams have already incorporated into their workflow, it creates operational friction and, in some cases, internal resistance. Employees who have gained real productivity from these tools don't see cuts as a technical decision, but as a step backwards.

Moreover, the narrative of a "slowdown" can lead to overly conservative decisions. Cutting back without analysing which uses generate value and which don't is as reckless as deploying without measuring. The reasonable objective is to reach operational maturity in which AI spending is predictable, proportional, and linked to concrete results—not to eliminate it out of budgetary inertia.

For teams like ours, working on custom integrations with Claude, this moment sends a clear signal: the conversation with clients has shifted from "How do we deploy this?" to "How do we manage this sustainably?". It's a change of question that actually indicates maturity.

Editorial perspective: Companies reviewing their AI invoices is not bad news; it's a sign that the sector is moving beyond the uncontrolled testing phase. The challenge now is ensuring that review is analytical, not reactive.