“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

• Tell us, does your company have a separate budget for AI tools? If so, what is it in 2026?

Yes, the company has not just one, but a whole system of budgets for AI tools divided by purpose. Monthly operating expenses for infrastructure and inference alone amount to hundreds of thousands of dollars. And this is without taking into account budgets for R&D and development of its own platforms. For the company, AI remains both one of the main areas of growth and investment.

The key approach of the company is that budgets are targeted. We don’t just give out "money for tokens", but allocate funds to specific areas:

• Budget for training for certifications, for example, over 5,500 of our specialists are already certified by Claude Certification from Anthropic and the token limit is designed specifically for training needs and the opportunity to pass the certification.
• Operating budget for routine tasks: working with documents, creating assistants, building a second brain, automating processes. Available to all employees.
• Project budgets, when AI is part of the delivery, are calculated individually for each project, depending on its specifics.
• R&D and infrastructure budget for the development of our own AI products (AI/Run, DIAL, CodeMie) and infrastructure support. These funds are divided into three main categories of expenses:
• Infrastructure and LLM-inference for our internal products.
• Subscriptions to market-leading AI tools (GitHub Copilot, Cursor, Microsoft Copilot, and others).
• API consumption (token-based pay) for complex agent workloads.

This approach with targeted budgets allows us to flexibly manage costs and invest where it has the greatest impact.

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

The average number will not be representative. Costs vary dramatically depending on the role, project, and type of tasks. Costs range from tens to several thousand dollars per month per employee. What determines costs: Product complexity — a greenfield project with minimal context or a legacy system with millions of lines of code. Typically, more context increases costs with each subsequent increment.

Development stage — active development of new functionality consumes more tokens than support.

AI Harness — how effectively the harness is built for working with AI.

Semantic layer — the amount of context that models need to process and its quality.

The type of tasks — architectural discovery — can cost thousands of dollars, and routine automation — tens.

There are developers who are satisfied with less than $100 per month, although they have the opportunity to use more. There are projects where the costs for one specialist reach several thousand dollars. Both cases are the norm, just different contexts of use.

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

All of the above. The AI ​​market for development has gone through several stages. Initially, the main criterion was models: "this one is better than the previous one", "ours is more accurate than the competitors'". Over time, the focus shifted from models to agent solutions. GitHub Copilot, Cursor, etc. Now the focus has shifted to AI Harness — a wrapper around agents for coding and skills. Modern implementations of high-quality AI Harness, which the industry calls AI Factory, are usually "ambidextrous" — they work equally well with market leaders.

We partner with OpenAI, Anthropic, Microsoft, Cursor, and other platforms. We use the entire stack because different clients have different input conditions due to partnerships, compliance requirements, or technical reasons. That’s why we focus not on the tool, but on the AI ​​Harness that we build around it. The tool can be changed, the architecture and approach remain the same. Modern harnesses are built so that tools can be easily replaced when needed. The main value is the semantic layer, pipelines, and quality gates.

— Do you see direct savings from using AI? For example, if a certain task used to take two days to complete, but now it takes a few hours.

Definitely yes. But it is important to understand that AI has different effectiveness for different types of tasks. We have a huge library of cases where AI has reduced costs by several times. There are tasks where the effect is smaller, only a few percent. It is important to understand that for business it is not so much about saving on a separate, isolated task, but more on the overall end-to-end process — from receiving requirements to releasing a finished product. After all, we think in terms of total cost — Total Cost of Ownership (TCO) for the customer.

The maximum savings (by times and tens of times) apply to deterministic tasks, where the requirements are clearly defined from the very beginning.

Moderate savings (in percentages) usually apply to new products where the final vision of the product is not yet formed and will change. When a customer comes with an idea, the process involves many iterations: discussion, concept creation, prototyping, feedback gathering, adjustments. The code development itself is only part of this process. AI can automate individual phases (for example, writing code for a prototype), but the overall process remains iterative and requires active human participation. There are savings here, although somewhat more moderate than in the previous example.

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

Yes, AI optimizes costs in all these areas, but the main result is not just reductions, but a fundamental restructuring of processes to achieve efficiency. The biggest savings in software development are our core business, and this is the industry that is currently being transformed the most by AI. Product support is also undergoing changes: automation of first-line support, generation of documentation, automatic analysis of logs and incidents — all this is now done by AI.

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

We don’t use superficial metrics like code written or team reduction. Our approach is based on three fundamental business indicators that directly impact the value of the product to the customer. This is a classic "priority triangle", where the customer chooses what is most important to them.

On top of TCO, we calculate efficiency gain. This is a percentage efficiency measure that shows how much AI can reduce this total cost.

Time to Market (or Lead Time). This is the time that passes from the moment a business idea is born ("let’s make this button") to the moment it appears to users in production. For some customers, speed is as important as price or quality.

Product Quality. There are clients for whom quality is an absolute priority. They don’t care how much it cost or how quickly we did it. What matters to them is that the product is flawless and reliable.

With the advent of the Mythos and Mythos-like models, Security and Vulnerability testing has reached a new level. For example, the open agent solution Visa Vulnerability Agentic Harness (VVAH) allows you to detect security issues that would have been extremely difficult and time-consuming to detect without AI.

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

Definitely yes. The company has developed a multi-layered security mechanism, especially regarding customer data.

Key rules:

• Using AI for client data always requires authorization from the client and we agree on the rules with each customer.
• It is prohibited to use AI without the client’s consent (this is stated in our security policies and mandatory training).
• There is a clear list of information that can never be transmitted to AI services.
• Technical protection: AI Harness has guard rails, i.e. automatic checks that scan data locally before sending it to the model. If sensitive information is detected, the request is simply not sent.

So it’s a multi-layered security model: policies, training, technical constraints, and monitoring. Data security is not an option, but a fundamental requirement for working with enterprise customers.

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

Trying to determine the average token spend per developer is like measuring the average temperature in a hospital: the number will be unrepresentative. Real costs are extremely variable and depend on a multitude of factors, making any average value meaningless.

Key factors affecting cost:

• Project and task complexity. Writing code from a clean slate (greenfield) with little initial context will be significantly cheaper than working on a large enterprise project with millions of lines of code.
• Infrastructure and Environment (AI Harness). The effectiveness of the so-called “harness,” i.e. the tools and processes built around the AI ​​model, directly impacts token consumption.
• Context (semantic layer). The amount and quality of context provided to the model to perform the task is one of the main drivers of value.
• Project stage. Development of new functionality, maintenance of existing code, or migration. Each of these stages has a different "price" in tokens.

To give you an idea of ​​the range, some developers spend less than $100 per month, and that’s more than enough for them. However, on complex projects or during the discovery phase, when architects are analyzing large amounts of data, with high workloads, especially in agent mode, the costs per specialist can reach several thousand dollars in a short period.

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

In an ideal world, there would be no difference, but reality is more complicated. Modern AI Harness works on the principle that existed before AI: there are tasks of different complexity — simple and complex. Traditionally, simple tasks are given to junior specialists, more complex ones — to senior ones. In the AI ​​world, this principle is preserved, but now agents often assess complexity, and in some cases do it more accurately than humans.

Junior specialists usually work with simpler tasks that consume fewer tokens. Seniors take on more complex tasks that require more tokens for context analysis, solution generation, and validation.

Since most AI factories involve a human-in-the-loop process, where a human makes decisions at different stages of the pipeline, the division remains natural: junior specialists start with simpler tasks and gradually move on to more complex ones, while seniors work on complex tasks, where their expertise is most valuable. So yes, seniors spend more tokens, but this is a consequence of the complexity of their tasks, not their status.

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

The largest consumption of tokens is by architects, especially in the discovery phase. When you need to do extensive research, build a semantic layer, analyze a large amount of data, or plan a migration, then the costs can be significant.

An architect working in a "team of agents" pattern can spend a lot of tokens at once.

This makes sense: architects solve the most complex problems with the greatest context. They analyze existing systems, design new ones, and evaluate solution options. Each such task requires processing large amounts of information, which directly affects token consumption.

Developers come in second, but their spending is more consistent and predictable. Marketers, analysts, and managers use AI for specific tasks, but the consumption volumes are much smaller due to the different nature of their work.

— 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?

Yes, we do. The company does not have a single universal limit; instead, we have a multi-layered system of target limits, each of which has its own purpose.

• Training and certification limits. This is the simplest type. For example, to prepare for the Claude Certified Architect (CCA) certification, an employee is allocated a limit sufficient to study the API, complete test tasks, and successfully pass the exam. It is limited, but clearly designed for a specific training goal.
• A general limit for all employees. This is a broader limit, not tied to a specific project. Its purpose is to give everyone the opportunity to automate routine tasks, create a personal knowledge base or "second brain", develop a prototype, or simply experiment with tools.
• Project limits. If AI is an integral part of a project, separate limits are set for it. Their size is determined individually and depends on the complexity of the project, the size of the code base, the quality of the configured AI-Harness, and the depth of the semantic layer.

The warning system works proactively: when 90% of usage is reached, a warning is sent. When 100% is reached, access is automatically blocked. But if the limit is exhausted, we have a mechanism for quick continuation. We understand that the workload is uneven: a new presale, an urgent task, the start of a project — all these are objective reasons for increasing the limits.

We also teach people token economics: how to optimize costs, avoid unnecessary usage, and structure requests correctly. It’s not about restrictions, but about conscious usage.

— Do you think spending on AI is justified?

Yes, this is the best investment the company has made and continues to make. We see a clear correlation between investing in AI and increasing the efficiency of our work. AI is one of the best investments you can make right now.

It is important to understand: token costs depend on context and complexity. The simpler the task, the fewer tokens it consumes. The more complex the task, the more. There are projects with many simple tasks, there are projects with complex tasks, and the price is always determined by the context. When we write from a blank slate and the context is on two pages, there will be one price. When we work with an enterprise system, where there are a million lines of code in a poly-repo, the price will be completely different. But even with maximum token costs, these engineers become 2-3 times more productive. The ROI is obvious: faster delivery of value to the client, higher code quality, less time on routine, more focus on complex tasks. The key factor for success is proper use with reliable monitoring and performance analytics.

Cryptocurrency has become the attention economy: why the most valuable asset is not the token, but the user’s attention

On the topic

Cryptocurrency has become the attention economy: why the most valuable asset is not the token, but the user’s attention

Tokens have become the new prey of hackers: how the Claude account theft scheme works

On the topic

Tokens have become the new prey of hackers: how the Claude account theft scheme works

Microsoft employee spent $28,000 on AI in one month. Now the company is imposing token limits

On the topic

Microsoft employee spent $28,000 on AI in one month. Now the company is imposing token limits

Read the country’s main IT news in our Telegram

On the topic

Read the country’s main IT news in our Telegram