Why isn’t anyone using the AI tool we rolled out?

Across industries, the pattern is the same: frontline employees’ use of generative AI has plateaued at around half, while three-quarters of senior staff use the same tools several times a week ¹. A related piece of research puts a number on the consequence: 60% of companies are not generating any material value from their AI investment, despite the spend ². If you’ve bought the licences, run the training, and watched usage stay flat anyway, you are not managing a bad purchase. You are looking at a documented, widely repeated failure mode.

Deployed, adopted and valued are three different things

Procurement measures deployment: licences issued, logins created, a training module marked complete. None of that tells you whether anyone is actually using the tool to do their work differently, and it tells you even less about whether that use is creating value.

A tool is deployed when it exists on people’s desktops. It is adopted when people choose to use it without being told to, and keep choosing. It is valued only when that use changes an outcome worth tracking, not just a login count.

Where licences have been bought, deployment is close to universal. Regular use among frontline staff sits well below leadership usage ¹. Even where adoption looks healthy, many organisations stop short of the third bar: deploying a tool brings incremental improvement, but the larger, transformative gains belong to the smaller group of companies that go on to redesign how the work itself gets done ².

This is why training completion rates and dashboard logins are such poor evidence of success. They describe what was delivered to people, not what people decided to do with it. What people decide to do with a new system is a judgement call, which is exactly where trust comes in.

Why AI raises the stakes

Earlier generations of workplace software mostly automated a task: data entry, scheduling, reporting. You could mandate their use, because there wasn’t much discretion involved. AI tools are different. Using them well means deciding when to trust an output, when to override it, and how much of a task to hand over at all. That’s judgement, not compliance, and judgement can’t be ordered into existence.

That dependence on discretion collides with three things AI tools raise that older software rarely did: opacity (it’s not always clear why a system produced a given answer), ambiguous accountability (if the output is wrong, who answers for it?), and the fear of being watched or replaced by the thing you’re being asked to use. Put those three together and the low-cost response is to avoid real use, or to route around the tool entirely.

What people are actually asking

Underneath the flat usage numbers are a handful of unspoken questions that rarely surface in a rollout town hall.

• What does this mean for my role? A majority of workers report real concern about AI’s effect on their jobs, and a substantial share say they’re anxious specifically about AI replacing them ³. One academic review found that job-loss anxiety, though common, does not strongly predict whether people accept AI ⁴. It’s one concern among several rather than the dominant driver of resistance.
• Will this be used to judge me? This is where the evidence is strongest. Most employees already report some form of AI-related monitoring in their current job, and employees themselves say that more transparency about AI use would raise their trust in their employer ⁵. What people are reacting to is not knowing what the tool is being used to measure about them.
• Who’s accountable when it gets it wrong? This is a question people bring to a rollout, and comms plans are rarely built to answer it directly. Leaving it open is reasonable to expect to cost confidence, even without a tidy statistic attached to it.

None of these questions get resolved by better onboarding. They get resolved, or don’t, by what leadership actually says and does.

Resistance is information, not an obstacle

When adoption stalls, the instinct is to treat it as a behaviour problem: more comms, more training, a reminder in the all-hands. But low usage and workarounds are usually a signal about the tool, the process, or the unanswered questions above, not a signal about the people.

The workaround pattern is well documented. More than half of employees say they use AI tools even when not formally authorised to ⁶. The exact number matters less than what it tells you: people are already using AI to get their jobs done. They’re just not doing it through the system you bought, because the official route hasn’t earned enough trust to be worth using.

Sentiment is moving in the wrong direction, too. Mentions of AI in employee reviews have surged, and the balance of sentiment has shifted from overwhelmingly positive a few years ago to now tipped negative ⁷. Whatever is wrong with how organisations are introducing these tools, it isn’t fixing itself with time.

Engagement that works, and engagement theatre

A town hall, a slide deck, and a feedback form are not the same as involving people. The difference is whether anything changes as a result.

Theatre looks like: announcing the tool after it’s chosen, inviting questions at a briefing with no mechanism to act on them, and publishing an FAQ that avoids the job-impact question. Working engagement looks like: bringing end users in before the tool is picked, naming what the tool will and won’t be used to measure, and being able to point to a specific decision that changed because of feedback. The test is whether anyone can show what their input changed.

Practical moves

• Involve end users in defining the problem before choosing the tool. If the people doing the work didn’t help describe what needs fixing, the tool is answering a question they didn’t ask.
• Say plainly what the AI will and won’t be used for, including what it won’t measure about individual performance. Employees themselves say transparency about AI use would raise their trust in their employer ⁵.
• Create a real channel for concerns, and publish what changed because of it. A feedback form nobody can see the results of isn’t a channel.
• Pilot with real teams and let them set the rules for use, rather than handing down a usage policy written by people who won’t be using the tool day to day.
• Track use and perceived value, not just deployment metrics. Licences issued and logins recorded tell you almost nothing; what matters is whether people keep choosing the tool and whether it changes an outcome you’d otherwise measure.
• Have leaders use the tools openly, including their limits. A leader who won’t use the system, or who presents it as flawless, tells everyone else the real answer: that use is optional, or that admitting limits is unsafe.

Trust is the infrastructure

Trust is the infrastructure any tool runs on, not an afterthought to roll out once the technology is in place. Capability gets you deployment. Trust gets you adoption, and only sustained adoption, directed at real outcomes, gets you value.

Where in your current rollout are people being informed rather than involved?