
Many AI seats go quiet after the first few weeks
Use of paid AI seats peaks soon after launch, then settles lower. In a large Copilot trial, a worker's employer predicted their use better than their own work habits did.
Key takeaways
- In a 66-firm Copilot trial, weekly use peaked at 55% of licensed workers and settled just under 40% from week 121.
- Over the trial's last three months, average weekly use ranged from 6.3% of licensed workers at one firm to 70% at another1.
- In Gartner's 2024 survey, 57% of IT leaders said user engagement with Copilot declines quickly after it's implemented5.
- Gallup finds employees who strongly agree their manager supports AI use are 1.7 times as likely to be frequent users9.
Paid AI seats get the most use in the first few weeks. In a randomized trial Microsoft ran with 66 firms, weekly use of Microsoft 365 Copilot peaked at 55% of licensed workers and settled just under 40% from week 121. Firm averages over the last three months ranged from 6.3% to 70%1. Surveys and government evaluations find more use where managers expect it and back it, though none tested that at random.
What happens to AI adoption after rollout
Between September 2023 and October 2024, Microsoft worked with 66 large firms that each put at least 100 workers into a lottery for Copilot licences, and tracked 7,137 knowledge workers through its own telemetry1. A worker counted as a user in a given week only if they took an action in Copilot1.
More than 90% of licensed workers tried Copilot at least once in six months1. Weekly use peaked at 55% in the first weeks, then slid until it settled just under 40% from week 121. In months 4 to 6, 5% used it every week and 20% didn't use it at all1.
Weekly Copilot use in a six-month trial, as a share of licensed workers
Months 4 to 6: average weekly use at each firm
55% at the peak
In the first five weeks, 55% of licensed workers used Copilot in a given week. People were trying the tool and some firms ran introductory training. Over six months, more than 90% tried it at least once.
Dillon and colleagues, NBER working paper 33795. 66 firms, 7,137 workers. The dashed line joins the two published points; its shape isn't reported.
These were large firms that joined a study during Copilot's early rollout, so a smaller firm buying today may see different numbers. UK and Australian government pilots gave most licences to volunteers and measured use in their own ways, so their figures can't be compared directly.
Gave about 70% of licences to volunteers and 30% to randomly chosen staff, and tracked use in Microsoft's dashboard.
In an average week, 64% of licence holders used Copilot at least once. Daily use averaged 30% on working days.2
Gave Copilot to volunteer agencies and staff, and surveyed them before, during and after.
Of 811 people who answered the final survey's frequency question, a third used Copilot daily and 21% a few times a month.3
Gave Copilot to staff who applied, and surveyed them and their managers.
Use was lower than the department expected. Most used Copilot two to three times a week or less.4
Why employees stop using AI
Light users in Australia's final survey said they'd stopped after a poor first experience, or because checking and editing the output took longer than doing the work3. Others lacked confidence and couldn't find time to train, and focus groups said they often forgot Copilot was built into their apps3.
At Australia's Treasury, staff expected more than the product delivered, and some were discouraged and gave up4. They also struggled to fit learning the tool into their workload4.
of IT leaders said employees struggle to fit Copilot into their daily routines5
said user engagement with Copilot declines quickly after it's implemented5
said users need frequent engagement and education when Copilot is introduced5
Gartner surveyed 132 IT leaders in June 2024, about half at companies with 10,000 or more staff5. Training on each person's own tasks is part of the fix, and why training beats tools covers what kind works.
AI licence waste: what an idle seat costs
Microsoft lists Copilot Business at US$21 a user a month on an annual plan, before a first-year discount that runs through 20266. Its terms give buyers 7 days to cancel for a prorated refund, so after that an unused seat still costs US$252 over the year at list price6. Australia's Treasury estimated a mid-level staff member would need to move about 13 minutes a week to higher-value work to cover the licence, which assumes the seat gets used4.
What a year of quiet seats costs
Set your seats and price, then pick how many get used in a typical week.
- Quiet seats in a typical week
- 24
- All seats, a year
- C$13,920
- Cost of each used seat, a month
- C$74
Estimates from your inputs. The C$29 default is Microsoft's US$21 Copilot Business list price at about C$1.38 to the US dollar, so use your own quote. Use shares come from the 66-firm Copilot trial.
Few pilots grow into full rollouts. In Gartner's June 2024 survey, 60% of IT leaders had started Copilot pilots and 6% had finished one and planned a large rollout5. In May 2025, Gartner analyst Dan Wilson said 5% of organizations in its 2025 survey that had finished pilots were moving to a larger deployment7.
Taking seats back is awkward, since 90% of IT leaders in the 2024 survey said employees would fight to keep their access5. BCI, a Canadian public-sector investment manager, rolled Copilot out to 300 employees and has since rotated the licences across more than 570 users, according to a Microsoft customer story8.
What keeps AI use going: managers who back it and AI champions
Knowing a worker's firm predicted their use in the field trial better than all the researchers' other measures combined1. Industry explained only 2.9% of the variation1. The authors suspect differences in management practices or training that their data couldn't capture1.
Gallup finds employees who strongly agree their manager actively supports their team's AI use are 1.7 times as likely to use it a few times a week or more, but only 36% strongly agree9. In McKinsey's 2026 survey, the roughly 6% it counts as AI high performers were twice as likely to say senior leaders show commitment to AI and that they measure its impact10.
In Australian agencies where leaders expected Copilot use, focus groups reported better uptake, and one participant said use rose after the chief executive showed staff how he used it11. The evaluators recommended naming AI champions to show colleagues the benefits12. The UK pilot found that team and line manager attitudes appeared to shape use2.
What went with steadier use
Tick what your firm has in place. Hover or focus an item to see the evidence.
Managers back it out loudSurvey, Gallup; UK Department for Business and Trade
Staff who strongly agree their manager supports their team's AI use are 1.7 times as likely to use it a few times a week or more. In the UK pilot, team and line manager attitudes appeared to shape use.
Drawn from the government evaluations and surveys cited in this article. Each is a correlation or an interview finding, so the count is a checklist and makes no forecast.
Microsoft surveyed 1,300 of its Copilot users, and most said at 11 weeks that the tool had improved their work13. Those are the same weeks in which trial use fell from its peak to where it settled, so the first three months seem to decide whether the habit forms1.
AI adoption metrics to check every week
Microsoft's Copilot usage report counts a user as active if they tried a Copilot feature in the chosen window of 7, 28, 90 or 180 days, and admins can export each user's last activity date14. Gartner's Dan Wilson said the default counts one use in 28 days, adding, "I don't really call that adoption"7. He suggested one use a week as a floor and one a day to make a seat worth its cost7.
OpenAI's admin guide for ChatGPT Enterprise tracks a similar funnel, from seats bought to weekly active users, and suggests sharing the numbers with AI champions and department leads15. A small firm can run on four numbers:
- Weekly active rate by team: the share of paid seats used in the past 7 days.
- Quiet seats: anyone with no activity in the past 4 weeks, from the last-activity export.
- Active days per person, where the report shows them, to tell a habit from a one-off try.
- One task measure per team, such as the time a weekly report takes, before rollout and at week 12.
A 12-week AI rollout plan for a 10 to 500 person firm
- Before buying, name one owner who'll work on adoption every week for 12 weeks, and list three to five tasks per team for the tool.
- Buy seats for the people with those tasks and add the rest in phases, as Australia's Treasury recommended4. Check the cancellation window first, since Microsoft's annual terms allow 7 days6.
- In weeks 1 and 2, have each team practise its listed tasks on real documents.
- Name a champion in each team, as Australia's evaluators recommended12, and hold a 30-minute office hour every week where people bring real work.
- Ask managers to use the tool in front of their teams, since manager support goes with more frequent use9.
- Every Monday, have the owner check the 7-day active rate by team and follow up with the quietest team.
- At week 12, when trial use had settled, move seats idle for 4 weeks to people who've asked for one, and drop the rest at renewal1.
At list price, 40 Copilot Business seats cost US$10,080 a year before anyone's time for training6. To check whether the owner and the tasks are in place before you buy, try the readiness check. For assistants built on company documents, see what company AI brains did and needed first.
Questions owners ask
Why do employees stop using AI tools?
In Australia's government Copilot trial, light users said they'd had a poor first experience or that checking the output took longer than doing the work3. At Australia's Treasury, some staff expected more than the tool delivered and gave up4.
What is a good Copilot licence utilization rate?
In a 66-firm field trial, just under 40% of licensed workers used Copilot in an average week from week 12, and firm averages ranged from 6.3% to 70%1. Gartner's Dan Wilson suggests counting a seat as adopted only if it's used at least weekly7.
What does an AI champion do?
An AI champion shows colleagues practical uses for the tool and gathers what works. Australia's evaluators recommended naming them12. In a small firm, that can be one person per team who runs a weekly office hour and flags teams that have gone quiet.
How long does it take employees to adopt an AI tool?
In the 66-firm trial, weekly use peaked early and settled from week 121. Plan on about three months of weekly support and usage checks before deciding whether a seat is worth keeping.
Sources
- 1Shifting Work Patterns with Generative AI (Working Paper 33795). National Bureau of Economic Research, 2025-11.
- 2The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. UK Department for Business and Trade, 2025-08-28.
- 3Evaluation of the whole-of-government trial of Microsoft 365 Copilot: employee-related outcomes. Digital Transformation Agency (Australia), 2024-10.
- 4Evaluation of a trial of generative artificial intelligence (Copilot) in the Department of the Treasury: summary. Australian Centre for Evaluation, Australian Treasury, 2025-02.
- 5Microsoft 365 Copilot rollouts slowed by data security, ROI concerns. Computerworld (reporting a Gartner survey), 2024-09-27.
- 6Microsoft Copilot for Business: plans and pricing. Microsoft, 2026-10.
- 7Gartner: Microsoft Copilot hype offset by ROI and readiness realities. techpartner.news, 2025-06-26.
- 8BCI empowers employees with Microsoft 365 Copilot (customer story). Microsoft, 2024-12-03.
- 9Global Indicator: Artificial Intelligence. Gallup, 2026.
- 10The state of AI in 2026: On the road to ROI. McKinsey & Company, 2026-08-25.
- 11Evaluation of the whole-of-government trial of Microsoft 365 Copilot: whole-of-government adoption of generative AI. Digital Transformation Agency (Australia), 2024-10.
- 12Evaluation of the whole-of-government trial of Microsoft 365 Copilot: summary report, executive summary. Digital Transformation Agency (Australia), 2024-10.
- 13AI Data Drop: The 11-by-11 Tipping Point. Microsoft WorkLab, 2024-03-19.
- 14Microsoft Copilot usage report. Microsoft Learn, 2026-07-14.
- 15ChatGPT Enterprise workspace analytics guide. OpenAI Academy, 2026-09-17.






