Our employees are not using Copilot. Why?

When Copilot in your ERP is not being used, it is almost always for one of four reasons: the function does not hit a real pain point in daily business, the results on unclear data are wrong too often and cost trust, nobody has ever shown where the function sits in your own process, or it is unclear whether one is allowed to trust and follow the suggestions. All four are leadership and order questions. Another training session fixes none of them.

In order: the four reasons

1. No pain point: the activated function saves the individual no noticeable time. Then non-usage is rational. The mistake lies in the choice of function, not in the users' will.
2. Wrong results on unclear data: whoever had to correct a wrong suggestion three times stops asking. That is not an AI problem but a data problem wearing an AI face.
3. No place in the process: the function exists, but nobody has shown at which point of your own workflow it sits and what it replaces there. Tools without a place do not get used.
4. Unclear permission: may I accept the suggestion? Who is liable if it was wrong? As long as that is open, restraint is self-protection, not refusal.

What works: one lever per reason

Against the missing pain point, reversing the order helps: first measure the three most time-consuming transactions in the team, then activate the one function that targets exactly them. What Copilot concretely deliversdiffers considerably per workplace, and that is exactly why blanket rollouts fail.

Against lost trust only data work helps: duplicates, outdated items and ambiguous fields out, then the hit rate rises visibly. Against the missing place, one hour at the real workplace helps more than a slide-deck training: the function inside your own process, on your own documents. And against unclear permission a written rule helps: suggestions may be accepted, we check spot samples centrally, the named process owner is responsible.

How to measure whether it works

Not by login numbers, but by transactions: how many orders, documents or texts ran with AI involvement this month, and how many minutes does the transaction cost today compared with the measurement before the rollout? Two figures, once a month; the question needs no more controlling than that.

If the figures stay flat after a quarter, that is a finding, not a failure: then it was the wrong function or the data foundation does not hold yet. Both can be fixed, and both are cheaper than a year of paid, unused capability.

The special case of managers

Adoption starts at the top: if the management keeps ordering reports by shout-out instead of asking the system itself, the organisation learns that AI is a topic for others. The most effective single adoption lever is a manager who visibly works with the system in a meeting. That costs nothing and beats any campaign.