Why nobody uses the AI assistant after launch, and what to do about it
The assistant went live and a month later two people use it: the usual reasons (wrong answers in the first days, an extra step in the workflow, no visible sources) and what to do before and after launch. How to measure whether it is really being used.
Short answer: people abandon an assistant in the first week for three reasons. It answered wrongly a few times. Asking it a question means opening a separate window. There is no way to see where the answer came from, so it has to be double-checked anyway. The fix is to launch on the narrow set of questions the assistant reliably answers, to put it where the work already happens, and to show a source under every answer. What to measure is questions per person per week and the share of answers flagged as wrong.
The first three answers
Someone asks the assistant three questions. If one answer is wrong, they conclude that everything will need checking and go back to the old way: ask a colleague, or dig through a folder. There is usually no second attempt.
So at launch the assistant should answer a narrow range of questions whose quality was measured in the pilot, and say “I don’t know” to the rest honestly. Broad coverage in the first week does more harm than narrow coverage. How to measure that quality beforehand is covered in piloting an AI project.
One extra step kills the habit
If a sales manager works inside the CRM and the assistant lives on a separate site, every question means switching, signing in and copying the context across. By the third day they stop doing it.
The assistant goes where the work already is: into the customer record in the CRM, into the work messenger, into the operator’s window. Then a question takes one action, and the context, which customer is open for instance, is picked up by the assistant itself.
Without a source, the answer is not trusted
“Delivery is in fourteen days”, with nothing attached, sends the manager off to check. At that point the assistant has only added work. The same answer with a link to the document or the record in the accounting system is verified in seconds, and after a week people stop verifying it.
Who to bring in first
Not the whole department. Two or three people who are curious about the tool anyway, and a manager who uses it openly. A fortnight later they have examples where it saved them time, and those examples convince the rest better than any training session.
The training itself should be short: which questions the assistant understands, how to flag a wrong answer, and where a question goes when it does not know.
What to measure
Questions per person per week. A total question count grows on the enthusiasts alone and hides the fact that nobody else is asking.
Share of returning users. How many of the people who asked something last week asked something this week.
Share of answers flagged wrong. If nobody flags anything, the likely explanation is that nobody is using it.
Questions with no answer. The log of questions the assistant answered “I don’t know” to is a ready-made list of what to add to the knowledge base.
Common questions
Should people be required to use the assistant? A mandate produces token questions for the sake of the report. Better to remove the obstacles: the place in the workflow, the quality, the sources.
How long before adoption picks up? If questions per person are not growing in the first two weeks, they will not start growing on their own. Read the log and find the reason.
What about the people who do not trust it? Show the source and let them flag a wrong answer. When someone sees their flags leading to fixes, trust arrives faster.
The rule
Launch the assistant on a narrow set of questions it answers reliably, inside the window where the work already happens, with a source under every answer. Widen the set once the first users are asking something every week.
Related reading
Facing a similar problem?
Tell us what you are building. We will walk through the architecture and give you a budget range in one call.