The case for AI experimentation over an AI mandate

When AI adoption stalls, the first instinct is almost always to mandate it: make the tools non-optional, tie them to performance reviews, and wait for usage to climb. I understand the instinct, and it still rarely works. People resist because of how the rollout reaches them. It gets dropped on the whole company within months of a vendor contract, with barely any training and no clear tie to the work in front of them. So they comply on paper and route around it in practice.

And the usage a mandate produces is thin. Gallup's Q4 2025 workforce study found only 26% of U.S. employees use AI at work even a few times a week, and nearly half say they never use it at all. A mandate can make people log in. It cannot make the tool useful to the job they actually do. That is exactly where structured AI experimentation, done well, beats the mandate outright.

It can fail in the other direction too. Left unsupported, experimentation slides into invisible, off-the-books use, and that is the failure people are right to worry about. There are two ways this goes wrong: the blunt mandate on one side, the unsupported free-for-all on the other. What I trust is the middle path, structured, supported experimentation, where the structure itself is what keeps it honest.

The strongest counterargument deserves a straight answer: without direction, nothing scales, and pure bottom-up experimentation produces a thousand pilots that never move the business. That's real. BCG's 2025 work describes a silicon ceiling, with only about half of frontline employees regularly using AI despite leadership pushing it, and McKinsey finds pilots proliferating while value stays marginal. Experimentation with no structure does stall.

But look closely at what that evidence supports. It makes the case for structure, and structure requires no coercion. An aimless free-for-all needs direction. Leadership names the first real use cases, teams experiment on those with coaching and reinforcement, and only proven wins get scaled. That's not a mandate. The mandate-versus-experiment framing is a false binary; the axis that actually predicts adoption is supported versus unsupported.

Experimentation pays off when it happens on the work people already do. An abstract sandbox teaches skills nobody carries back to the job. The OECD's 2025 review of the evidence found generative AI cutting task time by 15 to 30% in real-world settings, and the largest gains went to the least experienced workers.

That's the quiet case for letting whole teams experiment instead of anointing a few power users. The biggest lift goes to the people who were furthest behind, which is what it means to say AI is for everyone. So give a team one easy, high-value task to try AI on this week, coach them through it, and the win is immediate and theirs. Anchor it in role-specific use cases so each function experiments on work that actually fits its day.

When 78% of people are already using unapproved tools, it's fair to worry that encouraging experimentation only sanctions more shadow AI. Done with structure, it does the opposite. People reach for unsanctioned tools because they're trying to get real work done faster. One survey found 60% judged the security risk worth it just to hit a deadline. The motive is good. The governance is what's missing.

Structured experimentation closes that gap by bringing the appetite into the open: sanctioned tools, clear guardrails on what never gets pasted where, and a prompt library people actually share. That last part matters more than it sounds. A prompt that works for one person becomes a shortcut for the whole team, so a good find spreads instead of getting rediscovered one desk at a time.

That's governance through enablement. It turns invisible shadow use into visible, supported practice, and nobody has to build a monitoring dashboard to get there.

There's a reason your adoption numbers understate reality: people hide their AI use. They worry it looks like cheating, or like they can't do the job without help. Others hide it because they aren't sure they're using it right, and they would rather stay quiet than get it wrong where the whole team can see. And as long as that fear is in the room, experimentation stays private and shallow, and the learning never spreads.

What changes that is psychological safety. Leaders have to make it genuinely safe to try an unfamiliar tool and to get it wrong in front of peers. Nobody gets this right on the first try, and no one was ever supposed to. That's a cultural choice, and it's the difference between a team that quietly experiments alone and one that learns out loud and gets better together.

All of this lands or dies with the manager. The failure pattern is the manager who announces the tool and never mentions it again. Worse is the one who treats adoption as a compliance scoreboard and watches who's using what.

The pattern that works is the manager as coach. Model curiosity, ask the team what they tried this week, celebrate a real win, and help the person who's stuck instead of flagging the person who isn't. Name a task, reinforce the behavior, and scale what's proven. That's how AI adoption actually takes hold across a team, and it's why a named champion with manager backing moves a group faster than any directive. Support beats surveillance, every time.