Written by —
%20Large.jpeg)
Written by —
A company decides to invest in AI. Teams receive Copilot licenses, development teams get coding assistants, and the first agentic solutions are deployed quickly: a ticket-handling agent for customer service, proposal automation for sales. Leadership’s message is clear: experiment boldly—this is our AI strategy.
Six months later, the CFO opens the invoices. API costs have multiplied, licensing expenses have increased, and no one can clearly explain what the company has gained in return.
This pattern is becoming increasingly common. At Amazon, employees reportedly launched AI agents to perform meaningless tasks because token consumption had become a performance metric. Uber burned through its entire 2026 token budget in just four months, while Meta, Microsoft, and Walmart have all had to introduce restrictions to control rapidly growing AI costs. Agentic workflows can consume up to 100 times more tokens than traditional chat-based interactions. When AI adoption is encouraged without proper governance, costs escalate quickly.
The executive team asks the obvious question: What are we actually getting from all of this? No one has an answer. The next decision seems inevitable: cut spending.
“So how do we reduce the costs?”
The CFO’s response is straightforward: reduce licenses, cap the API budget, and freeze new AI initiatives until the situation is under control. Given the information available, it’s a perfectly rational reaction.
Meanwhile, someone on the engineering team raises a hand. We could significantly reduce costs without reducing usage. The customer service agent sends the entire ticket history to the model on every request. Summarizing the context would reduce token consumption dramatically. The proposal automation uses the most powerful, and most expensive, model for every step, even though a lightweight model would be sufficient for classification tasks. Some agents lack proper boundaries and repeatedly retrieve the same information because they have no stopping logic. Addressing these issues could significantly reduce costs without hurting quality. In fact, quality might even improve.
Both perspectives are valid, but neither solves the real problem. The CFO cuts costs blindly because there is no way to distinguish valuable AI usage from wasteful usage. The engineer optimises individual systems but has no visibility into whether those use cases are actually worth optimising in the first place.
The real challenge is not about reducing costs, but about creating visibility.
Santeri Salonen at Recordly’s office. Photo: Jami Ivanoff, 2026
The company builds monitoring capabilities. Now it can see which teams and use cases consume AI resources, how much they consume, and where those costs come from. But is visibility alone enough?
All AI costs still sit within the IT budget. The CIO is responsible for keeping spending under control and must justify every cost item to the executive team. Business units, on the other hand, have little incentive to optimize their AI usage because, from their perspective, AI is effectively a free internal service. This challenge is reflected more broadly as well. According to research by Writer, 97% of executives say they personally benefit from AI, but only 29% see measurable organizational returns. The benefits are real but they fail to accumulate at the organizational level because the operating model doesn’t connect costs with business value.
So when the question of reducing AI costs comes up again, both the CFO and the engineer have better answers. They can identify where the highest costs occur and focus optimisation efforts there through better context engineering and smarter model selection. But one person remains silent: the business leader. They don’t see their own AI costs, nor do they have to justify them. Uber’s COO summarized the challenge publicly: individual employees are becoming more productive, but that productivity isn’t translating into measurable business impact at the company level. That doesn’t necessarily mean AI isn’t delivering value. It may simply mean that no one owns both the costs and the benefits at the same time.
The real challenge isn’t just about visibility either, but about allocation.
The company makes a structural change. AI costs are allocated to the business units that actually use and benefit from the technology.
Now the head of customer service can see the cost of the AI agent within their own budget. They can evaluate the trade-off: the agent costs X per month but reduces handling time by Y percent. Perhaps the context should be compressed further. Perhaps unnecessary retrieval steps should be removed. They ask the engineering team to make those improvements because the savings now directly affect their own budget. The legal department deliberately chooses a more expensive model for contract analysis because the cost of making a mistake is far greater than the additional token cost. The cheapest model is not always the cheapest solution. Meanwhile, the sales team discovers that one of its automations delivers too little value and decides to simplify it on their own.
“So how do we reduce the costs?”
By now, the question has changed. No one is asking how to minimize token consumption anymore. Instead, the organization asks: Are we spending tokens in ways that create business value? That is an ROI question, not a cost question. The same optimization techniques, such as context engineering, model selection, and setting clear boundaries for AI agents, are still important. The difference is that they are now applied where they matter most, because the people making the decisions can see both the cost and the value.
Photo: Jami Ivanoff 2026
