Companies Overpay for AI Because of Employees' Bad Habits, Capgemini Poland Warns

One chat window left open from morning to night, in which an employee handles invoices, edits a marketing report, and translates meeting notes one after another, looks like convenience. In reality, it's one of the main reasons companies' AI bills grow faster than the actual number of tasks would suggest, warn experts at Capgemini Poland.

Ballooning Context Windows

The problem described by Capgemini Poland experts is known in the industry as a "ballooning context." Every language model has limited working memory, or a context window, that holds the entire conversation so far. If an employee talks to the bot for eight hours without closing the chat, the model has to reprocess the entire conversation history from the morning with every new message.

This isn't just a matter of speed. Language model providers bill usage in tokens, chunks of text that the system has to read and generate. The longer the conversation history, the more input tokens have to be processed with each new question, even if the question itself is short.

The model doesn't just become more expensive to maintain, it also starts to behave less reliably over time - Hubert Nafalski, Project Manager Insights & Data, Capgemini Poland

Why Answers Cost More Than Questions

The second mechanism experts point to concerns price asymmetry. Output tokens, the ones generated by the model in its response, are noticeably more expensive in providers' pricing than input tokens, which the user types as a question. That means every needlessly verbose AI answer, for example rewriting an entire document instead of pointing to just the changed line, hits the budget harder than a long question would.

In practice, this plays out when an employee asks the model to fix one paragraph in a multi-page report, and the system responds by regenerating the entire document from scratch, complete with an introduction, summary, and boilerplate nobody needed. That kind of answer costs far more than necessary, and the user still has to manually track down the part that actually changed.

Simple Habits, Real Savings

The recommendations from Capgemini Poland's experts are fundamentally simple. The first rule is to open a new chat window for each new task instead of running one continuous conversation throughout the workday. The second is to paste only short, necessary excerpts of documents into the model rather than entire files. The third is to phrase requests so the model returns only the changed fragments of text, without unnecessary introductions and summaries.

The fourth rule concerns organizing the workday itself: it's worth finishing one type of task, such as a financial analysis, before moving on to a completely different one, like preparing marketing material, and doing so in a separate, fresh chat window. Closing a session after each task limits the growth of the conversation history, and with it the number of tokens processed in subsequent questions.

The Scale of the Problem

According to Capgemini Poland, the sum of such seemingly minor habits, multiplied across hundreds or thousands of employees using AI tools every day, can generate costs running into hundreds of thousands of zlotys a year at the organizational level. That's money companies could save without cutting back on how much they use AI at all, simply by changing how employees conduct their conversations with the models.

For Polish companies that have been rolling out AI tools into daily work en masse in recent months, the problem of hidden token costs matters all the more because subscription and API usage bills grow along with the scale of deployment. Without internal guidelines on how to use chats, companies can end up paying far more than the actual value AI delivers would justify.

Experts stress that the problem doesn't lie in the technology itself but in a lack of basic user education. Employees often don't realize that a longer conversation history translates directly into a higher cost for every subsequent message, because from their perspective the chat interface looks identical whether the conversation has lasted five minutes or eight hours.

The solution Capgemini Poland proposes isn't restricting access to AI tools, but a short training session on basic chat "hygiene," the same way companies teach employees the rules of safe email use or internal systems. Changing a handful of habits, with no additional investment required, can deliver measurable savings within a single quarter.