AI and Billable Hours

AI and Billable Hours

EPAM sells senior engineering hours, so generative AI is the factor that separates the two readings of its stall: a tool that compresses billable hours and invites price concessions, or a wave that enlarges the software its clients build. Through mid-2026 the labor model is still expanding — about 56,600 delivery professionals and utilization near 77% — even as EPAM's own filings escalated AI from an execution risk to a demand risk it says has hurt its stock.

Why the model is exposed to AI at all

EPAM's economics are simple to state and that is exactly why AI matters. The company employs people, trains them, staffs them onto client projects, and bills for their time; salaries of delivery professionals are booked as cost of revenue whether or not those people are on a billable project at a given moment [1]. Revenue is, to a first approximation, headcount times utilization times a billing rate. Any technology that changes how many hours a piece of software takes to build, or what a client will pay per hour, lands directly on that identity.

That cuts two ways, and both are real. If a coding assistant lets a smaller team ship the same system, a fixed-scope program needs fewer billed hours, and clients may — in EPAM's own words — "seek other service providers or expect price concessions" once comparable work can be done less expensively [2]. If instead cheaper software means clients commission more of it, the pie grows and an engineering-led vendor captures more work. The case is most sensitive to which force dominates, and it is not yet settled by the numbers.

Management's language escalated, on the record

The most concrete evidence that this is a live risk is that EPAM's own risk disclosure changed. The escalation is datable across three consecutive 10-Ks.

Source: EPAM FY2023 10-K [3]; FY2024 10-K [4]; FY2025 10-K [5].

In the FY2023 filing, the AI risk was mostly about EPAM keeping up: if it were "slow to develop, adopt, and deploy generative AI technologies," its competitiveness against peers would suffer [6]. By the FY2025 filing, EPAM had added a distinct risk factor titled "Increased Adoption of AI-Based Software Tools May Reduce Demand for Our Services," warning that clients could "develop, customize, and maintain software solutions internally," or "replace traditional software with agentic AI, rather than purchasing our services" [7]. The same page contains the most direct admission in the file: "Increased competition, or the perception of increased competition, from new and non-traditional market participants like AI-based task-specific tools, has negatively impacted the price of our stock" [8]. Management is not disputing that AI sits behind the de-rating; it is confirming it.

What the filing labels as an AI risk still shows a cyclical demand pattern underneath. EPAM's FY2025 10-K now names AI as a risk that 'may reduce demand for our services' and 'negatively impacted the price of our stock', but the stall it points to had a cyclical shape — Business Information & Media fell 10.5% and Consumer, Retail & Travel 5.6% in FY2024 while Life Sciences grew 17.3% — and the only vertical behaving as AI substitution predicts, Software & Hi-Tech, is EPAM's once-largest and now among its softest at ~15% of FY2025 revenue. That vertical breakdown is developed in [Demand Mix].

The original stall was macro, not AI

It matters that the two shocks arrived in a specific order, because it bears on whether the damage is structural. EPAM's revenue growth broke well before AI displacement was a commercial force. Revenue fell 2.8% in 2023 and grew 0.8% in 2024, after +28% in 2022.

Source: derived from reported financials, EPAM FY2021–FY2025 10-Ks; consolidated statements of operations [9].

The June 2023 outlook cut names the cause plainly. EPAM reduced its Q2 and full-year guidance because clients "become even more cautious with spending specifically in the 'build' segment," with "pipeline conversions occurring at slower rates than previously assumed" [10]. The FY2025 10-K reaches the same conclusion looking back: growth "significantly slowed at times, particularly during 2023 and the first half of 2024, due to reduced client demand resulting primarily from uncertain macroeconomic conditions" [11]. The first leg of the stall, then, was a cyclical pullback in discretionary IT spend. The AI-displacement worry is the newer, additive overhang layered on top of a business that was already growing slowly — which is why untangling the two is the analytical task, not a footnote.

What the operating numbers show so far

If AI were already gutting billable demand, the plainest tell would be a shrinking, less-utilized delivery force. So far the opposite is visible: EPAM has kept adding engineers and kept them busy.

Sources: FY2025 10-K, human-capital disclosure [12]; Q4 2024 earnings call [13].

Delivery professionals rose from about 47,350 at the end of 2023 to 55,100 in 2024 and 56,600 in 2025; utilization of that force improved from 74.3% to 76.7% to 76.8% over the same years [14]. Into the latest reported quarter, Q1 2026, the headcount held at more than 56,500 delivery professionals with utilization at 77% [15].

Source: derived from reported financials and human-capital disclosure, FY2025 10-K [16].

The revenue-per-professional column looks alarming for 2024, but that dip is an artifact, not a signal: EPAM closed the NEORIS and First Derivative acquisitions late in 2024, adding "nearly 6,000 people combined" to the year-end count while those staff contributed almost no 2024 revenue [17]. Normalized in 2025, revenue per delivery professional recovered to roughly $96,000. There is no evidence in these lines of AI silently hollowing out billed hours — at least not yet.

The AI-linked revenue itself is real but small and, tellingly, still unquantified. Management says 75% of its top clients are engaged on GenAI initiatives [18] and that AI-native services are growing "double digits" sequentially, but the CFO conceded it is "somewhat challenging to quantify precisely" [19]. On economics, CEO Balazs Fejes was measured: AI projects are "not fundamentally different from non-AI projects," and profitability is "currently similar to non-AI," with only the prospect of "higher profitability in the future" [20]. The upside case is a hope with a pilot behind it, not yet a margin in the accounts.

The bull mechanism management leans on

EPAM's answer to the displacement fear is a specific claim: cheaper software leads to more software, and AI pushes clients toward custom builds rather than off-the-shelf products. Fejes argues that "as functional capabilities improve, we anticipate a greater inclination to build rather than buy" [21]. Through 2024 and 2025 the company recast itself around this thesis — an "AI-native" delivery model it brands "Delivery-as-Code" [22], built on proprietary frameworks (DIAL, EliteA, and an AI/Run stack) meant to let clients run larger transformation programs with EPAM's engineering at the core [23]. Coming out of the trough, organic constant-currency growth did re-accelerate, from about 1.4% in Q1 2025 toward a mid-single-digit pace later in the year [24]. That is consistent with a cyclical recovery beginning — but it is modest, and the 2025 headline growth of 15% was inflated by acquisitions and currency, not organic demand.

The industry is testing the same question

This is not an EPAM-specific experiment; the entire IT-services sector is running it, and peer commentary shows both mechanisms operating at once. On the efficiency side, Cognizant reports that roughly 30% of its internal code was already AI-generated in a recent quarter and expects "it could reach 50% in the years ahead" — a direct measure of how fast machine-written code is displacing hand-written hours [25]. Infosys frames the current cycle around clients prioritizing "cost reduction and operational efficiency," which is "driving vendor consolidation" [26]. Both point at pricing pressure.

On the other side, Grid Dynamics — a smaller engineering-led peer closer to EPAM's model — reports the pie-expansion effect directly: because "the cost of development [is] getting reduced" by AI, it sees "increased demand for our custom-built software," and it is pursuing "enhanced pricing with our AI offering" rather than discounting [27], [28]. The honest reading of the peer set is that the net effect is genuinely undecided across the industry — which is why no single filing settles it for EPAM.

What would decide it, and what to watch

The evidence available today leans against the hard structural-decline case: through mid-2026 EPAM's delivery force is larger and more fully utilized than before generative AI, AI-native work is growing rather than cannibalizing, and the sharpest part of the stall traces to a macro spending pause that predates AI displacement. The strongest fact against that read sits in EPAM's own 10-K — the company itself now warns that AI adoption "may reduce demand for our services" and concedes it has weighed on the stock [29], and the effect is early enough that the operating metrics have not yet had to reveal it.

What would move the read is measurable and near-term. Revenue per delivery professional is the cleanest signal: if it falls on organic, acquisition-adjusted numbers while utilization stays high, that is price and hour compression showing through. A second is whether AI-native revenue becomes a disclosed, growing line rather than a "double digits, hard to quantify" aside [30]. A third is the character of the current softness: in Q1 2026 management again blamed macro — clients "modestly delay decisions" amid "higher energy prices and global economic uncertainty" — and cut full-year organic constant-currency growth to a 2.5% to 5% range [31]. If deferred deals convert in the second half as promised, the cyclical reading holds; if they quietly disappear, the case that AI is shrinking the addressable work gets harder to dismiss.