Enterprise AI: why 10,000 requests can hide just 14 real users

An analyst from the tech industry described on Habr a pattern he observed at four companies of different sizes and industries in 2025: corporate reports about "AI transformation" systematically hide dismal real-world usage numbers.

A pattern that repeats at every company

All four cases followed the same script. A request came down from the top: "We need AI, everyone else already has it." Teams rushed to deploy a corporate chat on top of an LLM, usually in the company's brand colors. A month later, the board saw a slide: "The AI assistant has been used 10,000 times." The board was happy. A line about "AI transformation" showed up in the investor letter. The product team got a bonus.

The author gives concrete numbers from one of the cases he examined:

• 10,000 — total requests to the assistant during the reporting period
• 9,000 of them — one-off visits from users who never came back
• about 500 — mandatory testing by employees and demo sessions for leadership
• 14 — actual weekly active users one quarter after launch
• 4,000 — the company's total headcount

The report still said "10,000." The slide was shown to the board twice.

Why the "requests" metric says nothing about value

The author lays out three questions that separate real adoption from window dressing. First: how many users came back a second time? Second: who is still using the tool once a week a month after launch? Third: what exactly have these people stopped doing by hand?

In the real cases, the answers to all three were uncomfortable. A large share of the requests on the impressive slide came from people whom management had asked to "try it out," and from demo sessions for top executives. Users who came back on their own could be counted on one hand.

"The report still said '10,000.'

The slide was shown at board meetings two more times," the author notes.

The mechanism is simple: a request count is convenient for anyone who wants to show progress without proving value. It is easy to inflate with initial curiosity, mandatory onboarding, and demo traffic. Meanwhile, the pressure on teams keeps growing: competitors have "adopted AI," investors expect a line about transformation, and success is usually measured by the first obvious number — the request count.

How to tell real adoption from adoption on paper

The way the original task gets framed is telling: when the order from above is "we need AI like everyone else," nobody asks what problem the tool is supposed to solve or how to measure whether it solved it. Without an answer to that question, any rollout is doomed to be a showcase.

The real metric of AI adoption is changed employee behavior. Not "how many times was it used," but "what did people stop doing manually." If a user tried the assistant and came back the next week without anyone asking, that signals real value. If 14 people out of 4,000 show up weekly after a quarter, that is a failure, and calling it a transformation is inaccurate.

Companies that are actually changing their processes with AI and companies that simply launched a chat in their brand colors look identical in current reports. The difference will show up later — in productivity, competitive advantage, and the ability to cut operating costs.

What this means

Most corporate reports on "AI adoption" in 2025 record the fact that a tool was launched, not a real change in how work gets done. The true benchmark for successful adoption is weekly active users, 30-day retention, and the tasks employees have stopped doing by hand.