Your team isn't resisting AI. They're resisting being automated without a say.

There’s a conversation happening in a lot of leadership meetings right now. It goes roughly: “We’ve rolled out the AI tools, we’ve done the lunch-and-learn, and the team just… isn’t using them. How do we get buy-in?”

The framing is the problem. “How do we get buy-in” assumes the decision was right and the people are wrong. Nine times out of ten, what’s being read as resistance to AI is actually resistance to something much more reasonable: having your job redesigned by people who didn’t ask you, using tools you don’t trust, on a timeline you heard about in an all-hands.

That’s not a change-management failure to be massaged away. It’s information. And the numbers say it’s everywhere.

The numbers behind the discomfort

Four statistics, all recent, all pointing the same way.

In Australia, the National AI Centre’s adoption insights (December 2025 – February 2026) found that around 65% of non-adopting SMEs cite distrust of AI decision-making, or a preference for human control, as their blocker. Not cost. Not capability. Trust.

Gallup’s 2026 research found only 9% of employees feel “very comfortable” using AI at work. McKinsey found just 12% of US employees have actually integrated AI into their daily workflows, despite the tools being deployed nearly everywhere. And BCG found that 70% of companies have trained less than a quarter of their workforce on the AI tools they’ve bought.

Read those together and a picture emerges: organisations are deploying tools at scale, training almost nobody, and then describing the predictable result as “resistance”. BCG’s own 10-20-70 rule says successful AI deployment is 10% algorithms, 20% tech and data, and 70% people and process. Most rollouts spend in exactly the inverse proportions.

Resistance is rational

Put yourself in the seat of the person being asked to adopt. Someone bought a tool you weren’t consulted on. The vendor demo showed it doing a chunk of your job. The internal comms say “AI won’t replace you, it’ll empower you”, but nobody has actually told you which of your tasks it’s taking, who decides, or what happens to the judgement calls you currently make.

In that position, slow-walking adoption isn’t irrational. It’s the only lever you have. Your team’s hesitancy is them telling you, in the politest available language, that they don’t trust the process that produced this decision.

And here’s the part leadership tends to miss: the resisters usually know something true. The person who’s done the job for six years knows which edge cases will break the shiny demo. The bookkeeper knows which supplier always sends mangled invoices. Resistance is frequently your most accurate map of where the automation will actually fail. Treating it as an attitude problem means throwing away your best QA data.

The fix is involvement, not messaging

We run our own company (an AI CRM for hospitality, operating across four markets) on an agent fleet: finance close, customer onboarding, support triage, marketing production. So when we say the fix is structural, it’s not theory; it’s how we’ve had to do it ourselves, and what we now do with clients’ teams.

Three things, in order.

Pick the tasks with the team, not for them. The task-selection workshop is the single highest-leverage hour in any AI rollout. Ask people directly: which parts of your week are repetitive, annoying, and beneath your skills? Which parts require judgement you’d never hand over? People are remarkably honest about this. Almost everyone has a list of work they’d happily never do again, and almost nobody nominates the parts of their job they’re proud of. When the automation targets come from the team, the adoption problem mostly evaporates, because there’s nothing to adopt: they asked for it.

Make “agents take tasks, people keep judgement calls” true in the design, not the deck. Every company says some version of this line. Very few build it. The difference is human-in-the-loop gates that are structural rather than optional: the agent drafts, a named human approves; the agent flags, a human decides; nothing posts, sends, or commits externally without sign-off. In our own finance close, agents do the reconciliation legwork and every journal still gets human approval before it posts. Your team can see the gate. That visibility is what converts a slogan into trust, because the question your sceptics are actually asking is “where does my judgement still matter?”, and a gate is an answer you can point at.

Train on their workflows, not a curriculum. Generic AI training is now a zero-price market: the vendor academies give it away free, and it changes almost nothing, because the gap was never “what is a prompt”. The training that works takes the team’s actual Tuesday (this inbox, this report, this customer complaint) and works through it with the tools, live, with the person who owns the workflow driving. One real workflow beats ten hours of slideware. This is also where you find out, cheaply, that the automation design was wrong in three places. Better in week one than month six.

What we tell our clients’ teams in the first session

When we start an engagement, we ask for a session with the team (not just leadership) and we say roughly this:

“We’re not here to automate you. We’re here to automate the parts of your job you’d nominate yourself, and you’re about to nominate them. The judgement calls stay with you, and we’ll show you exactly where in the system that’s enforced, not just promise it. Nothing this system does will be invisible: you’ll be able to see what the agent did and why, and overrule it. And you’ll know more about where this breaks than we do, so when you think it’s wrong, say so. That’s not resistance, that’s the job.”

Then we do the task-selection exercise, and the room changes. Scepticism doesn’t disappear (it shouldn’t), but it converts from passive resistance into active specification. The sceptics become the people writing the edge-case list. Same energy, pointed at the work.

The honest caveat: this approach is slower at the start. A mandate ships faster than a workshop. But the mandate’s speed is an illusion: you ship in week two and spend six months with 12%-style usage numbers, which is to say you didn’t actually ship anything. Involvement costs you two weeks up front and pays it back in actual adoption.

The uncomfortable question for leadership

If your team is resisting, the useful question isn’t “how do we overcome the resistance”. It’s: what do they know, or suspect, that we haven’t addressed?

Usually it’s one of three things. They suspect the real goal is headcount reduction and nobody will say so (if it is, say so: people can handle honesty far better than they handle being managed). They suspect the tool was chosen for the demo, not the workflow (often correct). Or they suspect that “human in the loop” is a transition state on the way to “human out of the loop” (address it by putting the gates in writing: which decisions stay human, reviewed on what cadence).

Answer the real objection and the adoption problem becomes a training problem. Training problems are easy.

Do this next week

Run a one-hour task-selection session with your team. Two columns on a whiteboard: “work I’d happily hand to a machine tomorrow” and “decisions that should stay with a human”. No tools, no vendors, no slides: just the lists.

You’ll leave with a ranked automation backlog your team has pre-approved, a written record of where the human gates belong, and (if you read the second column carefully) a fairly precise map of what your people believe their actual value is. All three are worth more than the software licence you were about to buy.