Марія Бровінська AI Eng 30 September 2026, 09:00

"We don't count tokens, instead, we choose tools where counting them is not necessary. 95% of colleagues use Claude": how much Railsware spends on AI tools and how it calculates it

AI has quickly transformed from an experimental tool into a part of the daily work of IT teams. But it’s interesting to see how much it really costs businesses and whether companies have started to calculate AI costs as systematically as cloud services or software. dev.ua asked Anatoly Yevpak, Associate Director at Railsware, about budgets, the most popular tools, and the token economy within the company.

— Does your company have a separate budget for AI tools? If so, what is it in 2026?

It would be strange not to have a budget for an AI tool in 2026. However, it is not static and changes: new services appear, existing ones deteriorate or improve, someone revises pricing policy, offers new subscription plans or changes the availability of individual models. As of now, Railsware’s spending on AI tools (if we are talking about “pure” AI tools, not software with added AI functions) is about $15–20 thousand per month.

At certain points, for a short time, we had higher costs. Let’s say a new tool came out or we were testing several at the same time to choose the best one for our tasks. Usually, in the end, teams switch to a single tool or choose one as the main one and keep an additional one on the baseline for small tasks.

— How much does a company spend on average on AI services per employee per month?

The average costs for our structure and business model will not be indicative, as we have different roles, a number of our own products, projects for clients, etc. For example, for some specialists it may be $100 or less, while for others it may be $200-300 or more. In addition, in our service branch of the business, where engineers actively use AI, for some employees the tools are paid for by clients.

— What tools are currently the most popular in your teams — ChatGPT, Claude, GitHub Copilot, Cursor, Gemini, others?

Currently, the most popular is Claude — it is used by 95% of colleagues, including specialists on our service projects, where licenses are covered by customers. The reason for this popularity is not surprising: it is the best price/quality ratio, a large number of use cases, and general popularity in the tech industry. In addition, with reasonable use (proper work planning and maximum use of subscription limits), it is the most effective for every dollar paid. Yes, on the premium license on the Team plan, which almost everyone at Railsware has, the limits are quite enough so that in 95% of cases they do not limit the ability to work full-time throughout the week.

ChatGPT is still used by about 20% of specialists — mainly managers, products, and the HR team. But it is more of an auxiliary, rather than a primary AI tool. Previously, it was used more actively, but now many tasks have been partially or completely transferred to Claude.

GitHub Copilot, similarly, is more often a secondary tool. While 15% of colleagues used it last winter, now there are three times fewer of them.

Cursor was surprisingly popular at the beginning of the year — back then, there were normal limits and “fast” queries, and every fourth of us used it. Constantly changing pricing policy pushes us to switch to other tools. Now, at Railsware, one in ten colleagues uses it, half of them are managers who need it as a code base analysis tool, and the rest are engineers. When we see that engineers mainly use Anthropic models in their work, we switch them to the premium Claude — it is several times cheaper for similar results. Cursor, on the other hand, can be unreasonably expensive with active use. Let’s say, one query costs $10-20 — it’s not uncommon.

Gemini (along with Antigravity with Google AI Ultra subscription for a small portion of active users + standard enterprise plan for the rest) was popular while Google offered discounts and almost unlimited usage. However, only 5% of employees were really active (mostly managers who work with a lot of text) and used it together with Antigravity. After Google increased prices and introduced hard limits, the tool lost its meaning. Now almost no one uses it. What is offered “out of the box” in the enterprise package has very small limits that make it unsuitable for practical use.

— Do you see direct savings from using AI? For example, if a task that used to take two days to complete now takes a few hours. Give specific examples.

There is definitely a saving in terms of work done for every dollar spent on AI, but it varies greatly by role, project, workflow, etc. AI is best used in application prototype development. What used to require a whole team of engineers with a designer and product and 2-3 months of work can now be done in a week with a not-so-technical product.

The work of engineers has also accelerated significantly, sometimes several times, depending on the project and processes. Now everything is not about writing, but about checking the code, which is still much faster than writing it without AI.

In general, the introduction of AI has affected almost every role in the company. Almost everyone has tasks that used to take hours or days, but now can be done better in minutes, and also save time and nerves. Even product managers now find it much easier to understand technical issues if they can simply talk to AI about what interests them, in particular, when there is access to the code base. This improves understanding of the project, simplifies or partially automates the work on planning and describing the scope for engineers, etc. Designers can make clickable prototypes no longer in Figma and do not spend a lot of money and a lot of time on it, but assemble them from ready-made components right away with the help of AI.

— Has the company reduced its hiring, development, or product support costs thanks to AI?

The costs for these articles remained at about the same level. But, at the same time, everything started happening faster, and the intensity of work increased, as did its volume. Nevertheless, hiring has become more difficult due to the advent of AI: a lot of spam, fraud, generated applications for positions, etc. We continue to experiment, in particular, we started using an AI recruiter who conducts initial interviews with a candidate, and also introduced “AI literacy” criteria: certain levels of mastery of AI tools that we consider necessary for a particular position. It is important to note here that the fundamental position of Railsware remains that all important decisions should be made by people, and any output from AI requires review and critical assessment.

— How do you determine the ROI from using AI — time saved, amount of code written, release speed, team reduction, or other metrics?

We haven’t formally started calculating ROI yet. However, there are a bunch of metrics that we are already collecting and analyzing. All of them indicate higher productivity across all roles, without exception, which is felt in the results of work.

— Does the company have internal rules regarding what can and cannot be sent to AI services?

Railsware has general rules for using AI services and information and cybersecurity rules, in particular, not sharing passwords, private data, customer data, etc. with AI. In addition, all of our AI services are corporate, which means that they do not have the ability to learn from our data.

— What is the average cost of AI tokens/credits per developer?

We don’t count tokens, instead, we choose tools where they don’t need to be counted. Moreover, different models and their tokens often have a price difference of tens and hundreds of times. We also adjust processes (or skills) that allow us to reduce token usage and get the same result (for example, telling the model NOT to do certain things that are very expensive).

— Is there a difference between the costs of junior, middle, and senior specialists?

In our experience, the difference is more by role than by seniority. Senior management or team leaders tend to use AI much more intensively due to the specifics of their work.

— Who in the company uses AI the most — developers, designers, marketers, analysts, or managers?

It depends on the specific project. Most often these are managers and engineers, but designers are often among the most active users of AI.

— Do you set limits on the use of AI services or the number of tokens? If so, what are they? What happens when an employee exceeds them?

Instead of paying “on demand” at public API prices, we buy subscriptions. For example, if you use 90%+ of the Premium subscription limits on Claude, and then do the same amount of work through Cursor (and in fact, at large volumes above the subscription limits, this is the price of the public API), the cost can be 20–30 times higher.

In fact, in a month, with very active use of Cursor, you can burn through the equivalent of 1–2 years of Claude premium subscription (using identical Anthropic models). In addition, we allow colleagues to use the remaining limits for private projects (because they will “burn” anyway).

— Do you think spending on AI is justified?

100% definitely yes.

“AI is one of the best investments you can make right now.” How much does EPAM spend on AI - Senior Director of AI Engineering and Implementation EPAM Ukraine on budgets, tokens and ROI

“People who deny the threat of AI are being marginalized like those who denied the coronavirus.” Does artificial intelligence really threaten humanity?