The pattern repeats often enough to be predictable. An organization buys AI licences for the team. Someone runs an introductory session. Everyone agrees it is impressive. Three months later, a check of the usage dashboard shows a handful of enthusiasts and a long tail of people who logged in once.
The instinct is to blame the tool, or the training, or the people. It is usually none of those.
Nobody knows which of their tasks it is for
The introductory session demonstrates capability: summarise this, draft that, extract the other. Everyone nods, because the demonstration works.
Then they return to their actual inbox, which does not contain a neatly-formatted sample document. It contains their real work, which is messier and more specific, and the leap from "AI can summarise things" to "this specific recurring task of mine is a candidate" is a leap most people never make on their own.
Training that starts from the tool's features teaches the tool. Training that starts from the participant's own recurring tasks changes behaviour. The difference in adoption is not marginal.
The first bad answer ends the experiment
A new user asks something they already know the answer to — a sensible instinct — and gets a response that is subtly wrong. Not obviously, absurdly wrong. Plausibly, confidently wrong.
For someone whose professional credibility depends on accuracy, that single experience is often enough. They conclude the tool cannot be trusted and stop.
What they were never taught is that this is expected behaviour, not malfunction, and that the skill is in knowing which tasks tolerate it. Drafting something you will edit anyway tolerates error well. Producing a figure you will forward unchecked does not. Nobody explains the distinction, so users apply a single verdict to a tool that deserves a differentiated one.
The rules are unclear, so the cautious opt out
Ask most teams what they are permitted to paste into an AI tool and you will get uncertainty. Client names? Contract text? Anything from the case management system?
In the absence of a clear policy, conscientious people choose the safe default, which is not to use it. The most careful staff — often the most senior — opt out first. Adoption then looks like a generational or attitudinal problem when it is actually a policy vacuum.
A one-page document naming what is fine, what is forbidden, and who to ask removes more friction than any amount of enthusiasm.
There is no time to be bad at something new
Every genuinely useful tool costs more time than it saves for the first few weeks. That is true of a new CRM and it is true of AI.
If someone is at capacity, the rational choice is the method they are already fast at. Adoption does not fail from unwillingness; it fails because nobody was given room to be temporarily slower.
Teams that succeed usually protect a small, explicit amount of time for it. An hour a week, aimed at one recurring task, beats an all-day workshop with no follow-through.
Nobody senior visibly uses it
If leadership talks about AI but never shows their own use of it, the message staff receive is that this is a directive rather than a practice.
The most effective adoption signal is not a policy or a training budget. It is a manager saying, in a normal meeting, "I drafted the first version of this with AI and then fixed these three things." That single sentence does more than a workshop, because it models both the use and the verification.
What actually changes the outcome
None of these are technology problems. They are training, policy, and management problems wearing technology's clothing — which is why buying better licences never resolves them.
The organizations where adoption sticks tend to do the same handful of things: they train on real work rather than demonstrations, they teach judgment about when to trust output, they publish clear rules, they protect a little time, and their leaders use it in public.
That is a solvable list. It is just not the list most organizations start with.
