AI Adoption Is a Supervisor Problem, Not a Tooling Problem

If your team isn’t using the AI yet despite a team-wide or company-wide rollout, the problem is probably neither the tool nor your people, but that the rollout treated adoption as a tooling problem, whereas it is a training and leadership problem.

Below are the four mistakes I often see organizations make when introducing AI and how to avoid them.

Mistake 1: They shipped a tool and called it adoption

People differ in how eagerly they adopt a new technology. Roughly, there are four types:

• Self-starters: give them the tool, and they will figure out how to use it.
• Early users: they need the tool plus relevant, concrete examples of how to use it on their specific job.
• Late users: in addition to the above, they need to see that most people around them are using it successfully.
• Resistant users: they won’t adopt it unless doing so becomes a requirement.

Unless your hiring process was deliberately designed to select self-starters, your organization contains all four types, and the self-starters are a minority. A licenses-plus-announcement rollout serves only them. For everyone else, the rollout provides the tool but none of what their adoption actually requires: examples, social proof, and norms.

Generic training doesn’t fill the gap, because examples only convince when they are concrete and relevant to the person’s own work.

As a rule of thumb, if your training material is generic enough so that it could be given to a project manager, a salesperson, and an engineer alike, it is not specific and relevant enough to drive adoption.ShareCopy

Mistake 2: They treated supervisors as bystanders

People do not change how they work because the CEO asked them to in a company-wide email. They change when their direct supervisor visibly cares.

If the supervisor doesn’t use AI, the team won’t use it either. If the supervisor doesn’t care, daily, about whether their people use it, the team won’t care either.

This is not because supervisors are cheerleaders for corporate initiatives. It is because they are the most visible person in a team and they control the local standards of work: what counts as good enough, what gets reviewed, what gets repeated, and what quietly dies after the kickoff meeting.

Supervisors are the linchpin of any tool or process adoption, and a rollout that doesn’t actively enlist them is a rollout that quietly excludes most of the workforce (everyone except self-starters and, maybe, some early users).

Enlisting supervisors means two things: securing their genuine buy-in, and training them to coach their people, because you cannot assume they already know how to introduce a new way of working without it coming across as meddling.

Mistake 3: They trained everyone in group sessions

A one-hour training for fifty people looks efficient, with a low cost per head. However, it is mostly wasted, for the reason we saw in Mistake 1: adoption is driven by examples that are concrete and relevant to each person’s job, and no group session can be relevant to fifty people at once, unless these people have the same job and are exactly at the same level with regards to AI proficiency.

Large sessions can be useful for orientation, risk awareness, or shared vocabulary, but they do not generate adoption. One-on-one sessions (or sessions with two or three people who share the same role and proficiency) are far more effective: you sit with the person, take one of their actual tasks, and work through it with AI together. More importantly, you get the time to address their personal doubts and have them practice under your supervision, something that’s impossible to do with large groups. One-on-one sessions look expensive, but the time is not wasted, because each session tends to produce something a group session rarely does: an actual adopter. An hour that converts one person beats an hour that washes over fifty.

Mistake 4: They tried to automate entire jobs

The last mistake is aiming AI at a job’s core task, with full automation as the goal. But this is exactly where AI is weakest, where the cost of errors is highest, and where people’s professional identity puts up the strongest resistance.

Thankfully, jobs are not monoliths. They are ensembles of tasks, and tasks are ensembles of steps. For most jobs, the full core task cannot be delegated to AI yet, while the ancillary tasks around it (administrative work, meeting preparation, pre-mortems, reviewing drafts for errors and likely objections, etc.) are full of low-hanging fruit. The same goes for sub-parts of the core task that can be augmented rather than automated. Smaller bites are easier to make work, easier to verify, and easier to build confidence in.

For example: do not ask AI to “own customer support,” but use it to summarize long tickets, suggest reply drafts, flag missing information, and convert repeated issues into knowledge-base entries. Or do not ask it to “automate lead generation,” but instead ask yourself what parts of lead generation can be automated, such as researching background information for a given person, or helping you draft an email not from scratch (that will sound artificial) but fleshing it out from a few bullet points you provide. In each case, AI is not about replacing the professional judgment at the center of the job. It is about removing friction and tediousness.

Why mandates do not work

Notice that none of the four mistakes covered here is about tools or incentives. That is why responding to low adoption with mandates, dashboards, or bonuses doesn’t work well. At most, they can force visible usage, but they cannot create skilled usage.

Incentives can buy more of an existing behavior, but rarely create it from scratch. The latter requires direct interactions between supervisors and their subordinates.

The supervisor-led playbook

Here is what works instead, in three moves and one boundary.

1. Start simple, and start with supervisors. You may not need custom tools integrated with your databases to begin; a simpler path is giving people general-purpose AI assistants and training them to use those on their personal effectiveness, which captures a large share of the available gains at a fraction of the investment. Before anything reaches the teams, secure supervisors’ buy-in and train them to coach.
2. One use case per role, with an engineered first win. Pick a single ancillary task where AI demonstrably helps that role, prepare good prompts, and have the supervisor coach each person through the first use, one-on-one. There is nothing worse than giving someone a powerful tool with no demonstration: they will try it once, it won’t go particularly well, and they will lose interest. Frame the whole effort as you-and-them against an external enemy, such as the workload or the competition, never as “I need you to abandon your way of working.”
3. Build judgment, not just procedures. Once people use AI, the risk shifts to using it badly: trusting hallucinations, missing errors. The fastest fix is running hypotheticals: list five to ten scenarios and dilemmas, ask your team what they would do, and give feedback while explaining your reasoning. A session of thirty to sixty minutes can provide months’ worth of experience, including experience in spotting AI mistakes.
4. The boundary: do not automate before it works with manual prompting. Premature automation is expensive twice: first because you spend time integrating and polishing workflows that may not matter, and then because you lock people into brittle processes before they have developed the judgment to use them well. In this context, “automation” includes both technical automation and skill automation: agents, templates, custom GPTs, workflow integrations, playbooks, and anything else that turns a human judgment into a repeatable procedure. Do not start with these. Instead, first make the task work manually with prompts, examples, and supervision, and only afterward, codify what works. Otherwise you are not scaling a capability; you are crystallizing an untested guess.

Measure output, not usage

One warning about goals. If the goal you set (or the metric you track) is AI usage, you will get busywork usage: prompts performed for the dashboard, tools opened to be seen using them, and no improvement in results.

The goal is increased output: faster turnaround, fewer errors, less rework, more capacity for the work that matters. Usage is a means, and it follows naturally when the use cases are well chosen. This is also what eventually moves resistant users: not a metric telling them they must, but visible results making non-usage the locally strange choice.

The prize includes upskilling, not just automation

Everything above gets AI used. But the biggest return is not (just) the time saved by automating work; it is using AI to upskill your team.

AI can codify what your best people do (their standards, checklists, and playbooks) and make it available to everyone else as real-time feedback and quality control: drafts reviewed against your rubric, gaps flagged, improvements suggested, with the human still accountable for what ships. Done well, this is on-the-job training that brings everyone closer to the skill level of your best employee. The floor rises, and outcomes become consistent.

The organizations that win with AI won’t be the ones that automate the most tasks; they’ll be the ones that treat AI as a training-and-quality system. And a team adopted this way compounds: the skills and judgment your people build ride every improvement in the models.

Want to get better at using AI yourself?

This post is about leading a team to adoption. If you want to raise your own skills as an individual user, read Seven Ways to Get Better Results from LLMs at Work for office tasks, The Goldfish Student for giving AI the right context, and Contracts for getting consistent, verifiable outputs on recurring tasks.

Want help putting this into practice?

Getting AI genuinely adopted (and used as a training-and-quality system rather than a box-ticking exercise) is one of the most common problems leaders bring to me. If you rolled out AI and usage is low, or usage is high but results aren’t moving, that is the kind of problem I work on with leaders. Here is how to start.