The CEO of a PE-backed IT services firm should count AI as a margin gain only where delivery teams using the tool spent fewer hours per ticket than teams without it over the same quarter, with defects counted beside the hours, and on time-and-materials work only once the freed hours were billed to other work.
EPAM Systems, which its annual report, filed February 26, 2026, calls a global provider of digital engineering, cloud and AI-enabled transformation services, earns the majority of its revenues under time-and-materials contracts billed using hourly, daily or monthly rates. Among its risk disclosures: Our clients have asked, and may come to expect, that we use AI along with human delivery personnel to develop software for them at comparatively lower costs. On an hourly bill, an hour AI saves is an hour not invoiced unless someone bills it elsewhere.
In a February 2025 draft since published in Management Science, six researchers, four at Princeton, MIT and Wharton and two at the company that owns the coding assistant's maker, analysed randomized trials of that assistant at three companies; MIT's GenAI initiative funded two authors. At one, a professional services firm, 61.3 percent of 320 developers in several of its Southeast Asian offices got access from late July 2023 and the rest had none for four months. Among developers using the tool, in estimates that also weigh weeks after the rest got access, completed pull requests, meaning finished code sent for review, rose an estimated 17.94 percent, not statistically significant, while the share of code builds that succeeded fell an estimated 17.4 percent; the authors say this only weakly points to more trial-and-error coding. Pooled across the three trials and 4,867 developers, completed tasks rose an estimated 26.08 percent among developers using the tool, though each experiment is noisy and results vary across experiments.
In a preprint from METR, a research nonprofit funded by donations that has accepted no funding from AI companies but uses free tokens from several, 16 experienced open-source developers did 246 tasks on their own mature projects from February to June 2025, each task randomly allowed or barred from AI tools, mostly using one of those companies' models. They estimated afterwards that AI cut completion time 20 percent, while allowing AI actually increases completion time by 19%, and METR cautions readers against overgeneralizing from it.
In month one the head of delivery picks one fixed-price account and one time-and-materials account whose client contracts allow the tool, and splits each account's engineers at random into teams with the tool and teams without, with new tickets dealt between them at random. From the split to the quarter review both log hours per ticket in the ticket system used for billing, and the quality lead counts defects, meaning reopened tickets and failed builds, per team. At the quarter review the CFO sets hours saved on the fixed-price account against its fee and checks the utilization report for whether hours freed on the other account were billed or sat on the bench. An operating partner can ask each portfolio services company for the same comparison.
Stop after one quarter; fewer hours with more defects is cost moved into rework and stays out of the margin.
On February 24, 2026, METR said it believes developers are likely more sped up in early 2026 than its early 2025 estimates, yet called its follow-up trial an unreliable signal, mainly because more developers declined to take part as they do not wish to work without AI, and said its time measurements are unreliable for developers who use multiple AI agents at once.
