Tuesday, September 15$199late booking, seats still openBook a seat
AI Ninja AcademyLive AI classes
← All guides

My staff won't use the AI tools I paid for. What now?

It is usually a blocked process, not a bad attitude. Check three things before you blame anyone. The tool cannot reach the files the work actually lives in. Nobody changed the process, so the output has nowhere to go. And every use costs extra minutes to verify. Microsoft found 78% of AI users already bring their own tools to work.

By Uvin Vindula · Checked 2026-09-03 · 8 min read

First, find out whether they are using AI at all

Before you conclude your staff are refusing AI, check whether they are refusing your AI. These are completely different problems with opposite fixes.

The evidence says most people are already using something. Microsoft and LinkedIn's 2024 Work Trend Index found 75% of knowledge workers using AI at work (microsoft.com). Of those users, 78% brought their own tools, rising to 80% at small and medium-sized companies.

PagerDuty surveyed 1,250 office professionals at firms with $500 million or more in revenue in June 2026. It found 66% had used AI tools they believed were not permitted, and 88% had shared work information with public tools like ChatGPT (pagerduty.com).

KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. They found 58% of employees intentionally using AI at work, and 57% hiding that use and presenting AI work as their own (kpmg.com).

So the low seat usage on your dashboard is probably not abstinence. It is substitution. Your people picked a different tool, and did not tell you.

The three blockers, in order

Almost every unused license traces to one of these three. Work them in this order, because fixing the second before the first wastes everyone's afternoon.

One: the tool cannot reach the files. The work lives in a shared drive, an email thread, a job-management system, or a folder of PDFs. If the AI tool cannot see any of it, using it means copying and pasting. Copying and pasting is slower than typing for anything short.

Two: nobody changed the process. The AI produces a draft, but the form still needs filling by hand, the approval still happens in a meeting, and the file still gets renamed manually. The output has no home, so it is extra work rather than replacement work.

Three: each use costs extra minutes. Someone has to check it. The KPMG study found 66% of users relying on AI output without evaluating accuracy and 56% saying they had made mistakes at work because of AI (kpmg.com). Once a person has been burned, they add a verification step. That step is unpaid, invisible, and often longer than the task.

What it looks like versus what it is

What you see What it usually is The actual fix
"Nobody logs in" Tool cannot reach the shared drive or inbox Connect the data source, or pick a tool that already lives there
"They tried it once" Output had nowhere to go in the existing process Remove one manual step and make the AI output the input to it
"It's not accurate enough" No agreed checking step, so everyone invented their own Write a two-minute check and say it is the whole check
"They're too busy" Net time cost is positive, not negative Time one real task both ways with a stopwatch
"They don't like change" No explicit permission, so nobody wants to be first Say in writing which tasks are approved
"Only the young ones use it" Older staff use it privately and do not report it Ask what they already use, not whether they use anything
"They say it's not their job" Role and targets unchanged, so effort is uncompensated Change what you measure before you change the tool

Permission is the cheapest fix and almost nobody does it

Slack's Fall 2024 Workforce Index covered 17,372 desk workers across 15 countries. It found 45% did not have explicit permission to use AI, and 37% said their company had no AI policy at all (slack.com).

Earlier Slack research found desk workers at companies with established AI permissions were nearly six times as likely to have experimented with AI tools (slack.com). Trained workers were far likelier to report productivity gains.

The reason permission matters is not bureaucratic. It is reputational. In the 2024 Work Trend Index, 52% of people using AI at work were reluctant to admit using it for their most important tasks (microsoft.com). Another 53% worried that using it made them look replaceable. Slack found 48% would be uncomfortable telling their manager they had used AI for a common task (slack.com).

Nobody adopts a tool that makes them look lazy or dispensable. Fix the signal before you fix the software.

The seven-step diagnostic

Run this over one week. It costs you about three hours and usually finds the answer on day one.

  1. Ask three people what AI they already use. Not whether. What. Promise no consequences and mean it. You will get a list.
  2. Watch one person do one real task. Sit beside them. Do not coach. Note every copy, paste, rename, and switch between windows.
  3. Time the task twice. Once as they do it now. Once with the tool, including the checking. Write both numbers down.
  4. Find the file blocker. Ask where the source information lives, then try to get the tool to read it. If it takes more than two minutes, you found blocker one.
  5. Find the destination. Ask where the output has to end up. If a human still retypes it, you found blocker two.
  6. Write the checking step yourself. Two minutes, three items, on one line. Then say out loud that this is the entire required check.
  7. Change one measured thing. Remove a step, a form, or an approval. Adoption follows the process change, not the announcement.

Price the verification tax honestly

The reason step three matters is that people do arithmetic you never see. If the tool saves eight minutes of drafting and adds six minutes of checking, it saves two minutes. Nobody rearranges their day for two minutes.

Attention is already fragmented. Microsoft analyzed aggregated Microsoft 365 signals to 15 February 2025. It found employees interrupted every two minutes during core hours, roughly 275 times a day, handling 117 emails and 153 Teams messages a weekday (microsoft.com).

Into that, you added a tool requiring a new window, a new login, and a new judgement about whether the output is true. Of course it lost.

The fix is to attack the checking cost directly. Give the tool the source documents so its output can be verified by looking, not by remembering. Narrow the task so a wrong answer is obvious. Accept a smaller win that needs no checking over a bigger one that does.

Stop counting seats

License utilisation is the wrong number, and it is the one every vendor dashboard shows you. It measures logins, not saved work.

Measure one process instead. Pick the task from step two. Track the elapsed time from start to finished output, once a week, for six weeks. That is a number you can act on.

Then check the second number: how much of the work is happening outside your tools. Ask again in week six what people are using. If shadow use is up and your seats are flat, your tool is worse than the free one, and no amount of training fixes that.

If the tool is genuinely wrong, retire it cleanly

Sometimes the answer is that you bought badly. That is not a disaster, and pretending otherwise costs more than the license.

Say it plainly to the team: this one did not fit our work, here is what we learned, here is what we are trying next. Then approve, in writing, the tool people are already using, with a short list of what must never be pasted into it.

That last document is worth more than the subscription. The KPMG study found almost half of employees admitted using AI in ways that contravene company policy, including uploading sensitive company information into free public tools (kpmg.com). A rule people can follow beats a tool they will not open.

Where these figures come from

People also ask

A page of standing instructions that writes your weekly document for you

Rather do this with someone than read about it?

Automate reports and documents with AI is a live two-hour class built on exactly this problem. 25 seats, your own files, and you leave with a saved set of instructions that writes that document for you, the way you want it.

Read next

Can lawyers use ChatGPT and keep client confidentiality?

Yes. ABA Formal Opinion 512, issued 29 July 2024, permits generative AI use under the existing Model Rules.

Can my ChatGPT conversations be subpoenaed?

Yes, and they already have been. In New York Times v. OpenAI, a court ordered OpenAI to produce 20 million de-identified ChatGPT conversations.

How do I stop AI making things up in my documents?

You cannot turn it off, so you build a checking step instead. Feed the AI your own source files, make it quote and cite, then verify every name, number…