Six months after an AI enablement push at a company of 50 to 500 people, good looks like this: about half the trained group reached for an assistant on real work in the past five days, a third do it most days, and one or two workflows have quietly changed shape. Twelve months in, reach stops being the useful number. One output has to have moved, with a date on it.
Most AI adoption benchmarks you will find were measured somewhere else. The numbers filling the top of this search came from panels of companies with 1,000 or more employees, each with an AI owner and a change function. You have neither, which cuts both ways: your ceiling is lower on reach and higher on speed.
So here are the bands we underwrite for a mid-sized company, split by what they measure, with the research behind each one named.
Three numbers all get called adoption
Reach, depth and output change run on separate clocks, and most benchmark pages blend all three into one percentage.
Reach is how many people touched the thing. It moves in weeks, your admin console hands it over free, and a vendor will quote it.
Depth is whether use became a habit. It moves in months, and it is where programs quietly die, because a pilot spike looks like real adoption for six weeks.
Output change is whether a deliverable takes less time, costs less, or happens at all. It moves in quarters, and it is the only layer a CFO cares about.
A page reporting 70% weekly active and calling it adoption has measured logins. Keep the three apart, and read every benchmark by asking which layer it came from. Our guide to proving capability moved covers the method; this page is about the numbers you compare yourself against.
What good looks like at six months
Take a trained group of 30 to 80 people at a 150-person company, six months after hands-on sessions on their own work.
| What you count | Weak | On track | Strong | The comparison behind it |
|---|---|---|---|---|
| Used an assistant on real work in the last 5 days | Under 35% | 45–60% | Over 70% | A randomized trial saw nearly 40% of workers given access use it regularly across 6 months |
| Use it 3 or more days a week | Under 20% | 30–45% | Over 55% | National panel data put 9% of employed adults at every-work-day use in late 2024, against 23% weekly |
| Paid seats that convert to real use | Under 35% | 50–70% | Over 80% | Recon Analytics measured 35.8% for Copilot against 83.1% for ChatGPT across 150,000 subscribers |
| Named workflows where an assistant is now standard | None | 1 to 2 | 3 or more | Workflow redesign separates the roughly 6% of firms that count as AI high performers |
| People who can show a pattern they reuse weekly | Under 20% | 35–50% | Over 60% | No public benchmark exists. This band is ours, from what we see in rooms |
| Output you can point at | Nothing yet | One candidate named | One measured | Expect little here at month 6 and do not panic about it |
Two rows need a note.
The 45-to-60% band sits deliberately below the 70% figures on consultancy pages. The closest evidence is that randomized trial of 6,000 workers at 56 firms, where regular use across six months landed near 40%, at firms with Microsoft’s own enablement support attached. Treating 70% as a six-month expectation sets you up to call a working program a failure.
The last row gets misread the other way. At month six, no output has to have moved. Capacity appears before anyone redirects it, and a board paper demanding a changed metric at week 26 is reading the wrong clock. What has to exist is one recurring deliverable somebody agreed to time.
What has to be different at twelve months
Month 12 is not month 6 with bigger numbers. Other things come due.
| Question at month 12 | Failing | Good |
|---|---|---|
| Did the habit hold or decay? | Weekly use is below the month-6 figure | Flat or climbing, measured the same way |
| Is use deep or occasional? | Nobody clears a sustained-use bar | A quarter of the group used an assistant in 9 of the past 12 weeks |
| Has it left the early adopters? | Only ops and engineering | Weakest function within 15 points of the strongest |
| Did an output move? | Still no measured workflow | One dated before-and-after on a real deliverable |
| Does an internal example exist? | Training still comes from outside | A second group trained partly off your own patterns |
| Is the spend decided? | Seats auto-renewed, unexamined | Renew, renegotiate or stop, on evidence |
The decay row is the one nobody benchmarks and everybody experiences. Recon Analytics watched one assistant’s share of US paid subscribers fall from 18.8% to 11.5% in seven months. Habits re-form around whatever is easiest, so a figure measured once at month three says nothing about month twelve.
The breadth row is where mid-sized companies beat large ones. Finance sits three desks from operations, so a working pattern travels by lunchtime rather than by intranet post. A weakest function 30 points behind the strongest at month 12 is not a skills gap. It is a program that trained the willing and stopped.
The published benchmarks came from companies ten times your size
This is what makes most benchmark pages misleading rather than vague.
The 2026 Wharton and GBK Collective study is the best current read on whether enterprise AI work is paying off. It surveyed 1,416 people in July 2026 at US firms above 1,000 employees and $50 million in revenue, and found 80% of leaders reporting positive returns but only 28% calling them strong.
Useful, and not your comparison set. A 180-person company has no Chief AI Officer, which that study links to stronger reported returns, and runs no platform-standardization program either. Benchmark against those medians and you will conclude you are behind when you are smaller.
The firm-level data is more honest about where you sit. The Census Bureau puts AI use at 32% among US firms with 100 to 249 employees and 37% above 250, against about 20% nationally. A Federal Reserve note reconciling three surveys found roughly 18% of firms had adopted AI by the end of 2025 while about 41% of workers used it at work that November.
Far more workers use AI than firms report adopting it, because people bring their own tools whether or not anyone trained them. Your baseline at month zero was never zero, and a benchmark assuming it was overstates what your program achieved.
The department-level sets circulating on this topic need the same scrutiny. The most cited one puts engineering at 65 to 80% weekly use and operations at 40 to 55%, a gradient that matches what we see, but it discloses no sample size. Use its shape, not its decimal places.
Licensed is not active, and the gap is wide
Seat counts are the most flattering number in your stack, and the least informative.
Recon Analytics surveyed more than 150,000 US paid AI subscribers between July 2025 and January 2026 and measured how many people with workplace access convert into users: 83.1% for ChatGPT, 35.8% for Copilot, 34.0% for Gemini. Same companies, same policies, a 47-point spread.
That spread is not a verdict on product quality. It reflects where the assistant sits in the day and whether anyone showed people what to do with it. The rule for your benchmark: conversion below 35% at six months means the tool is provisioned and unused, and buying more seats worsens the ratio. Fix the denominator before the numerator, because the arithmetic on when that spend pays back stretches with every idle license.
One more trap in the same family. Survey figures on daily use run far above console figures, because people count the week they used an assistant twice as a week of use. Managers in the Wharton study reported 62% daily use; console-grounded conversion rates over the same period ran about half that. Both are real, and they measure self-image and behavior. Collect both and never round upward for a board.
Depth is already defined, and nobody uses the definition
Microsoft published sustained-use definitions in its analytics reference, and they beat anything the benchmark pages invented. A habitual user averaged 1 to 14 assistant actions a week and used it in at least 9 of the past 12 weeks; a power user averaged 15 or more on the same test.
Steal the test, whatever your stack. Twelve weeks, nine of them active. One enthusiastic person running bulk tasks cannot fake it, and it survives a quarter-end lull.
By month 12 we look for a quarter of the trained group clearing that bar and about one in ten at the heavier threshold. Below a tenth, the sessions produced interest rather than skill, and a short refresher clinic beats another full training day. The options for testing fluency across a workforce cover that measurement end.
The month-six signal that predicts month twelve
We did not expect this one when we started running these sessions. The month-six number that best predicts the month-twelve result is not weekly active users. It is whether anyone in the room can name the workflow that changed, in one sentence, without checking a dashboard.
“The Monday pipeline summary takes 40 minutes instead of two hours” is a program that will be alive next year. “Everyone’s using it, the team loves it” is one that will show 30% weekly use in March and a renewal nobody can defend. The first answer means someone attached the tool to a specific recurring piece of work. The second means the tool is a toy with good reviews.
The large studies find the same thing from the other direction. McKinsey’s 2026 survey of 1,719 respondents shows nearly three-quarters of its high performers fundamentally redesigning workflows, against one quarter of everyone else. Wharton and GBK found firms measuring AI against growth outcomes nearly three times as likely to report strong returns as those measuring efficiency alone, 35% against 13%. Both restate one point: adoption not attached to a named piece of work does not convert, which is the case made in why buying licenses is not the same as building capability.
So add the cheapest metric on this page to your month-six review. Count the people who can name their workflow. Three of forty is a warning. Fifteen of forty is a program with a second year.
Two numbers that mean stop
Benchmark pages are generous with targets and silent on failure. Two thresholds should end the spend.
Month 6, under 30% weekly use. Not 30% of headcount, 30% of the people you trained. The training did not transfer, and the honest move is a follow-up pass aimed at the work people are stuck on, not a second rollout. If a follow-up has already run and the number sits still, stop buying training.
Month 12, healthy use and no moved output. Harder to accept, because the dashboard looks fine. If 60% of the group uses an assistant weekly and not one deliverable is measurably different, the constraint was never skill. Nobody redirected the capacity. The fix is a management conversation about which backed-up work the freed hours should land on. Our piece on diagnosing an enablement push that did not work splits the two failures, which look alike from a distance and need opposite fixes.
Write both thresholds down before the program starts. A benchmark with no failure condition is a hope with a percentage attached.
Frequently asked questions
What is a good AI adoption rate after 6 months?
For the group you trained, 45% to 60% using an assistant on real work in a given week is on track, and 30% to 45% using it three or more days a week is a healthy habit figure. The closest evidence, a randomized trial across 56 firms, found regular use near 40% at six months, so the 70%-plus targets on consultancy pages are a strong result rather than a baseline.
Why is our Copilot usage so much lower than our ChatGPT usage?
That gap is the norm. Across 150,000 US paid subscribers, 83.1% of people with workplace ChatGPT access converted into users against 35.8% for Copilot. Benchmark each tool against its own conversion rate rather than against the other.
Should I benchmark against companies my own size or against the leaders?
Your own size for levels, the leaders for shape. Large-enterprise studies describe firms with an AI owner and a change function you do not have, so their medians sit at the wrong level. What transfers is the pattern: the gap between leaders and the rest is workflow redesign, and that is available to a 200-person company faster than to a 20,000-person one.
How many of our licenses should be active before we buy more seats?
At least half, and ideally 70%, among people who have been trained and had a month to form a habit. Growing seats while conversion sits under 35% worsens the ratio and hands finance a number that argues against the program at renewal.
What should change between month 6 and month 12?
Depth and evidence, not reach. By month 12 expect a quarter of the trained group clearing a sustained-use bar of nine active weeks out of twelve, the weakest function within 15 points of the strongest, and one dated before-and-after on a real deliverable. Reach can stay flat. Those three cannot.
Do people use AI as often as they say they do?
No, so plan for the gap. Managers in the 2026 Wharton and GBK study reported 62% daily use, while console-grounded conversion rates over the same period ran about half that. Self-report tells you about confidence, console data about behavior, and the lower figure goes in front of the board.
When should we stop spending on AI enablement?
At month six if fewer than 30% of the people you trained use an assistant weekly and a follow-up session has already failed to move it, and at month 12 if use is healthy but no output has measurably changed. The first is a learning failure, the second a management one, and only the first responds to more training.
How do I turn an adoption benchmark into a number my CFO accepts?
Pair the usage figure with one timed workflow and employer hourly cost, and show the assumptions separately so they can be argued with. A single confident percentage invites someone to find the input you buried, so our walkthrough of a finance-ready model sets them out line by line.
Key takeaways
- At six months, 45% to 60% weekly use among the people you trained is on track. Treat 70% as a strong outcome, not a target: the best randomized evidence lands near 40%.
- Reach, depth and output change are three numbers on three clocks. A benchmark quoted without saying which it measured is noise.
- Nearly every published figure comes from firms of 1,000-plus employees with an AI owner. Use their shape, not their levels.
- Seats flatter you. Conversion from access to real use ran from 34% to 83% across major assistants in one study.
- By month 12, reach can stay flat. Depth, breadth beyond the early adopters, and one dated output change cannot.
- Count how many trained people can name the workflow that changed. It costs nothing and predicts the second year better than any dashboard.
If you want these bands checked against your own group, tooling and the work you need to move, book a free AI readiness call. We will say plainly whether your numbers point to a learning problem, a management problem, or a program already working better than the benchmarks suggest.
Arthur Wandzel
