Study of 120,620 workers shows peak AI gains at moderate use, not maximum

ActivTrak's Productivity Lab conducted a three-quarter study from Q4 2025 to Q2 2026 tracking the same 120,620 employees across 1,009 organizations to measure how AI actually changes work behavior. The study categorized AI usage into three stages reflecting true operational progression rather than deployment metrics.

At Stage 1 (Research Assistance), 27% of employees used AI like a search engine to answer questions and summarize information. At Stage 2 (Task Execution), 14% used AI to draft content, generate ideas, and complete routine tasks they then validate and finalize — the stage where AI eliminates repetitive work, such as allowing a sales representative to generate a quote from five systems with a single prompt instead of manually compiling information. Only 2% of employees reached Stage 3 (Workflow Integration), where AI becomes an integral part of day-to-day workflows. Overall, 43% of employees studied used AI in any form.

The productivity data showed a striking pattern: metrics rise as employees move from little or no AI use to regular, task-level adoption, with healthy utilization peaking at 75%. However, once AI becomes embedded in workflows, healthy utilization drops about 5 percentage points — to levels statistically indistinguishable from employees who barely use AI. This finding contradicts traditional AI maturity models, which measure an organization's overall progress toward deeper AI adoption and tend to reinforce the perception that "most" usage is best.

The CEO identified two specific operational risks from pursuing maximum adoption without visibility into actual work impact. Organizations may lose sight of runaway costs: more mature usage means more powerful models, more tokens, and more infrastructure, and if the task does not require it, organizations are spending money they could invest elsewhere. At ActivTrak itself, when operations noticed Anthropic costs rising, they discovered employees routinely using the newest, most powerful model to rewrite customer emails — a task that did not require that level of sophistication. The organization then created an internal reference to help employees match the right model to the right task. Second, organizations may unwittingly foster operational disconnect, with employees sprinting ahead to generate sophisticated AI workflows that optimize individual tasks but without improving broader processes. If the workflow has not been redesigned around business goals, the result is more output, faster, but no real business value.

The research also revealed a durability finding with significant implications: 82% of employees who adopted AI kept using it, and once people move past casual use, they continue to use it quarter after quarter. Almost no one who goes deep ever comes back. This means adoption is a durable change, and the level of adoption a leader pushes their team toward is the level they will likely stick with. Driving everyone to the deepest tier may lock them into usage where productivity gains stall out and costs exceed benefits. The CEO's prescription is that leaders must define maturity targets intentionally, taking into account the makeup of their company, the type of work being done, and the goals being pursued — not adopting a single universal model. A lean AI-native company may need most people functioning alongside embedded AI workflows, whereas an established business may find AI task assistance sufficient competitive advantage with less operational disruption. The challenge ahead is not only how to increase the 2% who integrate AI into workflows, but also how to coach the 27% of novice users to reach task assistance fluency.