The promise of AI agents is seductive: automate repetitive tasks, reduce headcount, and operate 24/7 without human error or HR overhead. Startups and enterprises alike are rushing to replace internal roles with language model-driven agents—customer support, operations, content writing, even junior analysts. But for many, the AI dream is turning into a financial nightmare.
As usage grows, so do hidden costs. And in many cases, companies are discovering that AI agents can cost more than the staff they were meant to replace.
The New Reality of AI Economics
We are now seeing a wave of companies waking up to a harsh truth: LLMs and agent-based automations are not cheap. While introductory plans for tools like OpenAI’s ChatGPT or Claude from Anthropic may seem accessible—£20 to £300 per month—real-world deployment tells a different story.
Behind the scenes, most AI agents run on token-based pricing. Vendors charge by the number of tokens (chunks of words) processed during each interaction. Input tokens are counted when a prompt is sent, and output tokens when the model replies. These micro-charges add up quickly—especially in complex workflows where agents call each other, reference databases, or generate multi-paragraph outputs.
A Simple Scenario That Becomes Expensive
Consider an AI system used to handle customer support emails. A single email might be:
- Ingested and summarized by one agent
- Searched against a knowledge base by another
- Drafted and rephrased by a third
- Sent back for final formatting by a fourth
Each of these steps may trigger a separate LLM call. What begins as one ticket becomes four model invocations—and potentially thousands of tokens.
Multiply this by hundreds or thousands of support requests per day, and monthly costs can reach tens of thousands of pounds, just to keep the system running. At that point, a few human support agents may look like a bargain.
The Cursor Pricing Fiasco
A perfect example of this phenomenon occurred in mid-2025 with Cursor, a developer-focused AI IDE. Cursor built its product around GPT-4, promising embedded AI support for writing, debugging, and improving code. Initially, its Pro plan was widely adopted at $20/month. But in June 2025, Cursor introduced a new Ultra plan at $200/month and quietly throttled the capabilities of existing Pro users.
Developers found themselves unable to complete basic tasks without being told they had exceeded their token limits. Many were unaware that usage limits even existed. Complaints flooded Reddit and Twitter, with users accusing the company of stealthily degrading service mid-subscription.
After days of backlash, Cursor’s leadership issued a public apology and promised refunds to affected users. The CEO admitted that pricing changes had been poorly communicated, and the experience damaged trust in the brand.
But the lesson is bigger than one company: even reputable AI vendors can change pricing models abruptly, throttle access, or introduce steep new tiers with little warning.
Capgemini’s $25 Million Wake-Up Call
Enterprises are also getting burned. Capgemini CEO Aiman Ezzat recently revealed that the company had to cancel a generative AI chatbot project because the projected data usage would cost $25 million per year. The model was consuming so much data during context retention and generation that it exceeded budget forecasts by an order of magnitude.
Despite the appeal of automation, the cost-to-benefit ratio was unsustainable. Capgemini’s leadership realized they would be spending more to run the AI than to staff a traditional support team.
The Illusion of “Unlimited”
Many AI vendors advertise their tools with words like “unlimited,” “flat monthly rate,” or “enterprise-ready.” But these often mask usage caps, throttling mechanisms, or vague metering practices.
Some examples include:
- Token ceilings: Many tools offer a fixed number of monthly tokens, after which usage is slowed or disabled unless the user upgrades or pays per additional request.
- API overage fees: SaaS platforms that connect to OpenAI or Anthropic often pass through API costs, meaning that behind the scenes, you’re paying per use.
- Opaque usage dashboards: Users are often left guessing how many tokens their workflows consume, or when they will hit a usage wall.
This lack of clarity makes it difficult to budget or forecast costs. Worse, it encourages overuse—until the invoice arrives.
Vendor Lock-In: A Hidden Trap
AI vendors thrive on lock-in. The more your business builds internal processes, workflows, and tools around a specific LLM or agent platform, the harder it is to switch providers later. APIs differ slightly, data formats may not be portable, and prompt tuning often has to be re-engineered for new models.
The economic incentive is clear: once a company relies on a vendor for mission-critical tasks, it becomes difficult to leave—even if the costs escalate. Some call it the “new AWS bill,” referencing the way early cloud users underestimated their future cloud spend. Now, OpenAI, Anthropic, and others are the new infrastructure giants, and companies are once again caught in the trap.
How to Avoid Getting Burned
If you’re using AI agents in your business or planning to do so, here are five critical strategies to avoid excessive costs:
1. Audit Your Token Usage
Make sure you have visibility into token usage per task, per user, and per day. Treat it like you would a cloud spend dashboard. Monitor it regularly.
2. Set Spend Caps and Alerts
Use vendor tools—or build your own— to cap spending. Some APIs allow usage limits; if not, consider wrapping them with a proxy layer that monitors and restricts access.
3. Avoid Overengineering with Agents
Not every task needs a chain of five agents talking to each other. Test minimal viable prompts before deploying agent-based orchestration. Many companies overspend by building unnecessarily complex workflows.
4. Model Swap Readiness
Don’t get too tightly coupled to one model or vendor. Consider wrapper layers (like LangChain or OpenRouter) that allow you to switch LLMs with minimal disruption.
5. Evaluate Cost-to-Outcome
Just because something can be automated doesn’t mean it should be. Measure the actual business value each AI interaction delivers—and whether a human could do it more efficiently.
Final Thought: AI Should Augment, Not Bankrupt
The idea that AI will replace humans and cut costs is appealing—but in many real-world deployments, costs spiral instead of shrink. Companies that overcommit to agent automation without clear cost controls are learning the hard way that AI agents don’t come with fixed salaries or capped expenses.
From Cursor’s misstep to Capgemini’s multi-million-dollar scare, the message is clear: without proper planning, vendor discipline, and usage visibility, you risk replacing your staff—only to get handed a £200k invoice from your AI vendor.
And by then, it may be too late to turn back.
If you’d like help auditing your AI workflows or estimating LLM usage costs before rollout, get in touch. Better to plan for the bill—than be blindsided by it.