Robin Aro, Head of Services
Forrest Gump’s mother’s famous analogy about a box of chocolates fits surprisingly well when describing artificial intelligence: “AI is like a box of chocolates. You never know what you’re going to get.” The user gives a task but doesn’t really know what they’ll end up with—and especially not how many resources were used or how much the answer will cost.
In a business context, this quickly becomes a practical issue. When AI is integrated into applications, customer service, internal tools, or decision-making support via an API, every request begins to have a unit cost. AI is no longer just a response visible in the user interface, but a production capacity operating in the background.
In AI services, costs are essentially driven by tokens. Tokens are small units of text, instructions, data, and responses that make up the price of usage. The user sees a simple request: “summarize,” “analyze these documents,” or “find the best course of action.” Behind the scenes, however, the model may use a short or long chain of reasoning, read a large amount of background information, perform intermediate steps, and ultimately provide a concise answer. In some cases, the contents of the “chocolate box” are unsatisfactory, and the user has to ask again to finally achieve the desired outcome.
“Artificial intelligence is like a box of chocolates. You never know what you’re going to get.”
On the surface, the task may look the same, but the cost is not the same.
This becomes even more apparent when the actual production environment is taken into account. A company’s AI solution consists of more than just the model. It is surrounded by an application, integrations, access rights, data security, monitoring, logging, a potential search layer, a test environment, and maintenance. A single response may cost only a few cents, but as usage grows to thousands or millions of requests, the total bill becomes a matter for management.
We are likely to see price competition in the market, but not everywhere. Basic models and simple use cases will become cheaper. In contrast, the best reasoning models, long contexts, agent-based workflows, and demanding enterprise use cases may remain expensive or even become more expensive in terms of total cost. Even if the unit price of a token decreases, the volume of use and the complexity of tasks can increase the cost. We are already seeing a trend where the adoption of new models is either being postponed or pushed back by higher user fees or token pricing.
A potential IPO by OpenAI would intensify this pressure. In addition to growth, a publicly traded company is expected to demonstrate profitability, margins, and predictable business operations. This could mean aggressive price competition for entry-level products, but more precise pricing, usage limits, and company-specific agreements for premium capabilities.
That is why companies should not view AI as magic or as limitless automation. Instead, they should view it as a new production capacity with a unit cost.
The winners of the future don’t always use the most efficient model in every situation. They know how to choose the right model for the right task, limit the data set, set cost limits, utilize caching, measure the benefits, and track what a single customer request, analysis, or decision-making support actually costs. The worst offenders in the AI chocolate box are those use cases where the answer is always regenerated from scratch using vaguely modeled data, without leveraging the results of previous queries.
An AI “chocolate box” can hold a lot of value. But in business, you need to see the price tag before handing it out to everyone.
Yes, this text was written entirely by artificial intelligence, with the exception of a few corrections. Based on figures pulled from the API and Azure environment, its estimated cost is approximately 0.10 euros. Generating this text may have consumed approximately 0.5–3 Wh of electricity and an estimated 0.5–5 ml of water directly for cooling. The exact figure depends on the model used, the data center, the load, and how much computation is actually required to generate the response. The consumption estimates are based on the ChatGPT 5.4 model.
Behind every successful AI application lies a well-designed data platform where data is standardized, high-quality, and presented in a format that is understandable from a business perspective. Once a semantic model has been built for the data, AI can utilize it more reliably and effectively.
We help companies get more out of their data—at the right time, in the right format, and in a way that makes a real difference to their business. Feel free to ask us for more information.
