90% of Employees Say AI Training Is Useful—A Month Later, No One's Using It?

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

A study on M365 Copilot usage behavior reveals striking numbers: 90% of participants acknowledge formal training helped, yet 70% never watch training videos. What actually gets them using it is hands-on trial and error, plus discussions with colleagues.

Here’s an even more common scenario. A company rolls out AI literacy training for all employees. At course completion, everyone can articulate what generative AI can do. A month later, follow-up reveals actual adoption is minimal. The course quality wasn’t the problem—the problem lies outside the course itself.

Both sets of facts point to the same conclusion, and it’s not gentle: training doesn’t change behavior. Companies hoping to drive AI adoption through a single training program often buy themselves a month of psychological comfort.

Behavior Isn’t a Knowledge Problem, It’s a Structural Problem

Why do people attend, understand, then not use it? Because employee behavior is never determined by knowledge. It’s determined by three harder things: what’s written on the performance scorecard, what the manager does daily, and whether workflows have a place for AI.

Knowledge is the softest layer—you can pour it in within a few hours. Behavior gets stuck on the hardest layer: metrics haven’t changed, so time saved using AI doesn’t show up in KPIs; managers don’t use it themselves, so AI-assisted deliverables make employees look under-utilized; processes haven’t changed, so AI outputs still have to squeeze back into approval and reporting procedures designed for manual work. When structure doesn’t move, acquired knowledge naturally falls away, and people return to the path that was previously rewarded.

Background numbers show how widespread this is. In McKinsey’s global AI survey, 78% of organizations already use AI in at least one business function; another report on the future workforce shows 75% of U.S. Workers expect their jobs to change due to AI within five years, but only 45% have actually upskilled recently. That gap in between has never been filled by scheduling classes—it’s filled by moving learning from the classroom into the cracks of structure.

I have another observation that may sting: for many companies, the real buyer of AI training is management anxiety. They pay the budget and receive the reassurance that we did something. Judging whether a training falls into this category is simple—look at what else happens in the organization after training ends. If nothing happens, what they bought is peace of mind.

Change Has Three Depths, and Prices Vary Widely

Viewing AI capability building as a change initiative, you can divide it into three layers by depth of transformation. Each layer is more expensive than the last, and each gets closer to actual returns.

The first layer is foundational awareness. Help employees understand AI’s language, tools, and risks; remove doubts so they dare to try. Most companies invest most heavily here because it’s visible and measurable—you hold classes and get attendance records. But this layer only addresses whether people dare to touch it, not whether they use it. This explains why it’s easiest to produce beautiful activity photos and ugliest business results.

The second layer embeds AI into workflows and metrics. Contrast that company where training had no effect with another whose approach combined four things: integrating AI directly into workflows; managers leading by example and using it themselves; redesigning performance metrics to include continuous online learning time, scenario development, and efficiency KPIs; and building a peer-led support community. With the same course content, both awareness and application rose together. The difference wasn’t course quality—it was the supporting infrastructure. Notice this layer starts touching power and interests: changing metrics means redefining what good performance is; managers leading by example means publicly exposing their own learning curve. Most companies stop at the first layer, not because they don’t know the second is valuable, but because leadership’s resolve is only truly demanded starting at the second layer.

The third layer uses AI to rebuild the business itself. Develop AI application scenarios specific to your business domain that create competitive advantage, directly connecting capability building to business transformation. Few companies reach this layer, but performance gains happen precisely here—the first two layers are merely admission tickets.

Before Launching, Pass Four Self-Check Questions

If you can’t answer any of these four questions, hold off on the training budget. Training has never been the starting point of change—metrics and manager behavior are. Training just plants seeds in soil that’s already been loosened. Without loosened soil, scattering seeds only produces photo ops.

There’s a positive example worth referencing. When McKinsey rolled out its internal generative AI platform Lilli, it didn’t take the distribute-accounts-and-hold-classes route: leadership repeatedly told the transformation story, introduced new capabilities and processes alongside it, developed learning content, provided practical support, and ran capability building as a change initiative. It eventually reached 30,000 users and reduced the time needed to gain insights by 20%. Notice the structure of this case: story, process, support—nothing missing. Accounts and courses were actually the least important parts of the entire initiative.

When You Don’t Need This Much Effort

The boundaries of this approach need to be clear. For organizations with hundreds or thousands of people, where behavior is deeply locked by structure, it’s worth establishing a change initiative led by the CEO, not HR, because it touches metrics, processes, and manager behavior—all outside HR’s authority. A team of ten people needs nothing like this. The CEO using AI daily in their own work and demonstrating it in outputs is the most efficient change initiative—employees observe leadership behavior more deeply than they trust any course.

One more thing to be honest about: this approach assumes leadership truly wants to embrace change. If management’s calculation is merely using AI to cut costs and headcount, employees will smell it quickly, and no amount of upskilling design can substitute for genuine commitment. Framing learning as a long run the organization and you are taking together versus pushing fear onto employees—the output difference between these two narratives is far greater than any difference in course quality.

Next time before approving an AI training budget, do one five-minute thing: take out the company’s scorecard and read through it. If nothing on it rewards new ways of working, what this budget will likely buy is next month’s peace of mind.