5 decisions to make before you require your team to use AI

When an AI mandate becomes compliance theater

A PagerDuty survey that landed this week put a number on something I have watched play out across maybe forty deployments now: two-thirds of office workers have used AI on the job even when they believed it was against the rules. Sixty-six percent. Not because they are reckless. Because the tool was useful and the rule was vague.

Meanwhile the mandate lever is everywhere. Microsoft made internal AI use effectively non-optional and tied it to performance reviews. Meta baked “AI-driven impact” into every 2026 review regardless of role. Amazon’s review questions now ask how people used AI to move results. Accenture linked AI use to promotion eligibility. An AI Resume Builder survey reported by IT Brew back in February found that 58% of companies already require some set of employees to use AI tools.

So the instinct is understandable. Adoption is slow, the board is asking, and a mandate looks like leadership. Here is the thing I keep landing on: a mandate is a lever, not a strategy. Pull it before five specific decisions are made and the result is not adoption. It is the appearance of adoption, plus a quieter, riskier version of the real thing happening off to the side.

5 decisions to settle before mandating AI use

1. Decide what "use AI" actually means

   Pick the outcome, not the keystroke. "Use AI more" is the corporate equivalent of telling someone to "be more creative." Technically a goal. Functionally a shrug. When the instruction is an activity, people perform the activity: paste a prompt, screenshot the output, move on. When the instruction is a result, they reach for whatever gets them there. Write the requirement as something the team would want even if AI did not exist, like cutting first-draft time on customer replies in half.

2. Close the enablement gap before flipping the switch

   A Go1 survey covered by HR Dive in May found seven in ten professionals use AI weekly, but only 14% consider themselves advanced. That gap is the whole game. Nobody mandates their way out of a training gap. Before the policy ships, three things have to exist: approved tools people can actually reach, real training with time set aside for it, and a plain answer to "what am I allowed to put into this thing." Skip that and the mandate just pushes usage into the shadows.

3. Decide how real use gets measured

   Logins are not adoption. The fastest way to turn a requirement into theater is to count the wrong thing, because people optimize for whatever gets counted. Pick one workflow, write down a before number, and measure the after. If the only evidence is a dashboard of seat activity, a dashboard of seat activity is exactly what comes back. Measure work that changed, not buttons that got clicked.

4. Decide whether leadership goes first

   The same PagerDuty survey found 81% of workers believe AI rules get applied differently to leadership than to everyone else. And a Fortune piece in March, citing Nicholas Bloom's research, noted that many of the CEOs mandating AI are barely using it themselves. A requirement from someone who visibly does not do the thing is not a mandate. It is a memo. Go first, in public, and show the messy middle, not just the polished result.

5. Decide what happens to the holdouts

   This is the decision most companies skip until it is already a problem. Fortune's coverage of one 2,400-person survey reported that 29% of knowledge workers admit to quietly undermining their employer's AI push, rising to 44% among Gen Z, and that 60% of companies say they plan to let go of people who will not adopt. Lead with the second number and the first number grows. Separate "will not" from "cannot," because they need opposite responses: curiosity for the skeptic with a real objection, a clear and humane line for genuine refusal. Threats tend to manufacture the sabotage they are trying to prevent.

Why AI use mandates backfire

The pattern underneath all five decisions is simple. A mandate is a demand for compliance. Adoption is a change in behavior. Those are not the same thing, and confusing them is where most of the damage happens.

When a requirement arrives without the tooling, training, and clarity to back it, smart people do not stop using AI. They stop telling anyone. The PagerDuty numbers describe exactly that world: 43% have pasted work correspondence into public AI tools, 34% have entered customer data or information, and 31% have put financial or confidential documents into systems the company never approved. None of that shows up on the adoption dashboard. All of it shows up on the risk register, eventually.

There is a quieter cost too. An INTOO survey reported by HR Executive in May found that 25% of employees would not feel comfortable telling colleagues they had used AI at all. Push a mandate into a culture like that and the result is not more honesty about AI. It is less. Gallup’s February survey of more than 23,000 employees found 27% in AI-adopting organizations describe their workplace as having changed in disruptive ways to a large extent, against 17% in organizations that have not adopted. Disruption is not free, and a mandate spends that budget faster.

A mandate without enablement does not create adoption. It creates better-hidden adoption.

The companies getting this right are not softer on AI. If anything they are more demanding. They just aim the demand at outcomes and aim the support at people, instead of the other way around.

That 72% is the most useful sentence in the whole report. It means the people closest to the work already believe they know more about applying AI than the function writing the policy. A mandate ignores that belief. A good rollout uses it, by pulling the early adopters into shaping what “good” looks like instead of receiving it on a slide.

What a vague AI mandate actually produces

Here is what a vague mandate actually produces, according to the June PagerDuty data. None of these behaviors are what the policy intended, and all of them are what it caused.

Set that against the enablement side: seven in ten using AI weekly, but only 14% advanced. The distance between those two facts is the real target. Adoption is not a willingness problem at this point. It is a clarity, capability, and trust problem, and none of the three responds to an order.

Delay the mandate two weeks, settle five decisions first

If a mandate is already drafted, the move is not to scrap it. It is to delay the send by two weeks and spend that time on the five decisions above, in order. Define the outcome. Close the enablement gap. Fix what gets measured. Get leadership using it in public. Decide, in advance and with a clear head, how the company will treat the people who drag their feet.

The honest version of this is calmer than the headlines suggest. Adoption is already happening inside most companies, often faster than the official program. The job is less about forcing a reluctant workforce and more about giving an already-curious one permission, tools, and a definition of done. Do that and the mandate becomes almost unnecessary, which is usually the sign it would have worked.

The leaders I trust most on this are not the ones who moved first. They are the ones who made the team feel capable instead of cornered. That feeling is the actual adoption strategy. Everything else is paperwork.