Not long ago many companies treated AI tool subscriptions as a relatively cheap, predictable operating cost. A fixed monthly fee, access to advanced models, no need to dig deep into token usage or compute consumption. That model, however, is starting to come to an end.
The artificial intelligence (AI) market is entering a stage where providers are moving away, ever more visibly, from simple “all you can eat” plans. The reason is simple: the real cost of running AI models is high, and business usage is growing far faster than the original pricing models assumed.
The biggest shift is visible above all in agentic tools, meaning tools that do not just answer a single user question but carry out many operations automatically: they analyze code, run tests, generate follow-up queries, process documents and work practically non-stop.
For a human being, a usage limit comes naturally. A user may send a few dozen or a few hundred queries a day. An AI agent can generate thousands.
Spis treści
Subscriptions were priced for humans, then the machines started using them
This is the key problem for AI providers. Subscriptions were designed with interactive use in mind: someone sits at a computer, types a prompt, gets an answer, carries on working.
Meanwhile, more and more companies are integrating AI models into automated processes. This applies in particular to development teams, where AI tools can generate code, analyze repositories, propose fixes, run tests and execute tasks in a loop.
In that scenario a classic subscription stops making economic sense. The AI model starts to resemble an infrastructure service, similar to cloud computing. The more you consume, the more you pay.
That is why AI providers increasingly separate ordinary chat usage from automated or programmatic use of the models. The latter is moved into separate limits, credits, token-based billing or API fees.
Companies have to stop treating AI as a fixed subscription cost
For businesses this means one thing: AI budgets may become far less predictable.
Until now, rolling out AI tools often looked straightforward. A company bought licenses, gave employees access and assumed the monthly cost would stay more or less the same. With agentic tools that approach is risky.
If AI is wired deep into business processes, the cost can grow along with how hard those systems are working. And that intensity is often invisible to the board, the finance department or even the IT team, right up until the invoice arrives.
The situations that carry the most risk are those where:
- many teams use AI without any central oversight,
- AI tools are connected to repositories, documents or internal company systems,
- agents carry out tasks automatically,
- there is no monitoring of token usage and costs,
- the company builds critical processes on a single AI provider.
The hidden cost: not just the price per token, but also how tokens are counted
There is one more element worth noting, and it is often overlooked. The cost of using AI models depends not only on the official price per million tokens. It also depends on how a given model splits text into tokens.
A change of tokenizer can mean that the same text, document or prompt is counted as a larger number of tokens. Formally the price list may not change at all, but the actual cost of using the model goes up.
This matters especially with large data volumes: analyzing documentation, generating code, working with long contexts, automated tests or API integrations.
For a company it makes little difference whether the cost increase comes from a new price, a new limit or a different way of counting tokens. The effect is the same: a higher invoice.
AI is starting to look like cloud computing
We are seeing the same mechanism as with cloud services. At the start, getting in is easy, the cost seems low and the flexibility encourages fast rollouts. Only later does the need appear for cost control, architecture optimization and internal usage rules.
With AI this process can be even more dynamic, because generative models are very easy to deploy but harder to keep under control financially.
Companies should therefore treat AI not as an ordinary SaaS tool but as an infrastructure resource. Just as you monitor transfer, CPU, storage or database queries, you have to monitor tokens, model calls, the types of models used and the cost per process.
What should companies do right now?
The worst approach is waiting until the problem shows up on an invoice. If a company uses AI operationally, it should analyze several areas today.
First, you have to check the real usage of AI models. Knowing how many employees have access to a tool is not enough. You need to know which teams generate the largest costs, which processes use AI and whether that usage is interactive or automated.
Second, it is worth modeling costs in growth scenarios. A safe assumption is to check what happens to the budget if effective prices rise 2x, 5x or 10x. For many companies such a test may be uncomfortable, but it is better to run it before processes become fully dependent on a single provider.
Third, you should avoid full vendor lock-in. AI models should be integrated in a way that leaves room to switch providers, pick a cheaper model for simpler tasks or run part of the processes on open-source models.
Fourth, it pays to split tasks by cost and quality. Not every task needs the most expensive model. Simple classification, data extraction, technical summaries or routine operations can be handled by cheaper models, provided the system architecture is well designed.
AI still makes sense, but it requires cost control
This does not mean companies should give up on AI. Quite the opposite: well-implemented artificial intelligence can genuinely speed up team work, improve customer service, automate data analysis and lower operating costs.
The problem is not AI itself. The problem is rolling out tools uncritically, with no usage control, no cost architecture and no fallback plan.
The era of cheap, unlimited subscriptions is probably slowly coming to an end. Companies that treat AI like infrastructure will have an advantage. Companies that treat AI like just another simple SaaS plan may quickly run into unpredictable costs.
What it means for business
The coming months will show which billing models win out: simple subscriptions, token billing, credit packages or hybrid solutions. One thing is already clear: AI is no longer a “cheap add-on” to office work.
For companies this means a more deliberate approach to rollouts is now necessary. You have to measure usage, control costs, design integrations so that the provider can be swapped, and match models to the task at hand.
Artificial intelligence will become an ever more important part of company systems. But like any infrastructure, it has to be designed responsibly: technically, organizationally and… financially. If you are not sure how to optimize it, get in touch with us and we will sort it out together.
