Why nobody uses your AI tool (and why another training session will not help)
The licences are bought. There was a kickoff, a training session, maybe a second one. And when you look at the usage figures now, you see a handful of enthusiasts and a long tail of people who tried it once.
The usual explanation is resistance to change. That explanation is almost always wrong, and it is expensive, because it leads to another training session that will not work.
What actually happened
Somewhere in the first few weeks, someone asked a question with money attached. The maximum discount for a particular customer. Whether a shipment could go without prepayment. Whether something was still under warranty.
They got an answer. The answer was wrong. They only found out when a colleague said: no, that works differently here.
From that moment it is over. Not just for them. They mention it at the coffee machine, and within a week the whole department knows you cannot rely on it.
This is a studied phenomenon, not a character trait
Behavioural science calls this algorithm aversion. Research by Dietvorst, Simmons and Massey, published in the Journal of Experimental Psychology: General, shows across five studies something that seems illogical at first: people lose trust in an automated system much faster than in a person who makes exactly the same mistake. And they keep avoiding the system afterwards, even after seeing that it performs better than the person on average.
A colleague who is wrong once is still a colleague. A system that is wrong once is unreliable from then on.
That has an uncomfortable implication for your rollout: your AI investment does not fail gradually. It fails the moment someone is visibly wrong because of an answer from the tool. And that trust does not come back when you improve the system later.
And meanwhile, nobody checks
This is the most bitter part of the problem. Before that one visible incident, nobody was checking anything either.
Parasuraman and Manzey published in Human Factors a review of decades of research into what is called automation bias: people working with an automated aid under workload stop verifying its output. Two findings matter here. It affects experienced people as much as inexperienced ones. And it cannot be fixed with training or an instruction to stay alert.
Put those two side by side and you have the whole life cycle of a failed AI rollout:
- Nobody checks, so errors quietly pile up.
- At some point one surfaces, usually at a customer or on an invoice.
- Trust disappears in one go, and does not come back.
Why the answer was wrong
Rarely because of the model. Almost always because of what sits underneath.
Ask an AI tool what the maximum discount is for a particular customer, and it looks for something resembling an answer. It finds a message from the sales lead last week, an expired price list, a deck from last year's sales day and a sales handbook from 2021. All four true at some point. None of them carries any authority.
It picks one and says it with complete confidence. Not because it is lying, but because nobody told it which of the four applies, who decided that, and when.
What does have to happen
The order most companies follow is: roll out the tool, then train, then hope for adoption. That order is backwards.
What has to come first is this: gather the agreements that decide what is and is not allowed from the people who know them, have them confirmed by whoever is accountable, and record them with an owner, a date and a source.
This is not a documentation project and it is not a wiki. The difference lies in three things:
- Every rule has a name. Not "the department", but a person responsible for it being right.
- Every rule has a date, and a moment when it will be reviewed again. Without a review date everything decays, and decay is exactly how the last wiki died.
- Contradictions are not quietly resolved. When two sources contradict each other, a person decides. That decision is recorded, including who made it.
Once that is in place, something fundamental changes about the answer your people get. Not only is it right; it also shows who approved it and when. And an answer with a name under it is treated differently from an answer out of nowhere.
How to tell whether this applies to you
Three questions. Answer them honestly.
- If your most experienced colleague does not turn up tomorrow, does anyone else know what was agreed with your largest customer?
- Does your team use AI tools for work questions?
- Can you point out within thirty seconds who approved a given exception, and when?
Answer the three questions.
If the answer to the first and third is no, and to the second is yes, then low adoption is not a surprise but a predictable consequence. And the fix is not another training session.
Want to know where you stand?
Half an hour with two of us. Afterwards you get a single page: the agreements most at risk at your company, and who should own them. You keep that page even if we go no further.
Book the check →