How to Measure AI Adoption: 8 Metrics and a Scorecard

Why bother? Because adoption is easy to assume and hard to see. Statistics Canada found that 19.2% of Canadian businesses used AI to produce goods or deliver services in the 12 months before its spring 2026 survey, up from 6.1% two years earlier, and that 32.0% of those AI-using businesses had trained existing employees on it (Statistics Canada, June 11, 2026). Plenty of teams have paid for seats and run a session. Far fewer can say, three months later, what changed.

What "AI adoption" should mean for your team

Before choosing numbers, agree on a definition you could explain to the whole team in one sentence. A useful one:

AI is adopted when people use the approved tool, on a work account, for named tasks, repeatedly, and the checked result is at least as good as before.

Every part of that sentence becomes a metric. "Approved tool, work account" is the safety part. "Named tasks, repeatedly" is the habit part. "Checked result at least as good" is the quality part. If you only measure the first half, you'll reward people for opening the tool, not for doing better work.

Two rules before you start:

Measure teams, not individuals, by default. Per-person data is useful for offering help, not for ranking people. Microsoft 365 usage reports, including Copilot's, hide usernames by default "to help companies support local privacy laws", and only a global administrator can turn names on (Microsoft Learn, activity reports).
Tell people what you track and why. In Ontario, employers with 25 or more employees on January 1 must have a written electronic monitoring policy saying whether, how and why they monitor employees electronically (Ontario, Your guide to the ESA). If you'll look at per-user AI usage reports, make sure your policy covers it. Elsewhere it's still good practice.

Try it freeWant to see a session first? The free AI Hour is online on the first Friday of every month. Next: Fri, Nov 6, 12 PM ET.

The 8 metrics

1. Active-user rate

What it is: people who used the approved AI tool for at least one deliberate action in the last 28 days, divided by people who have access.

Where it comes from: your admin console. In Microsoft 365, the Copilot usage report shows enabled users, active users and the active-users rate over 7, 28, 90 or 180 days; opening the Copilot pane without submitting a prompt doesn't count as active (Microsoft Learn). In Google Workspace, Gemini reports show active Gemini users as a share of users whose licences include Gemini, over 7, 28 or 90 days (Google Workspace Admin Help). In ChatGPT Business, owners, admins and Analytics Viewers can view workspace analytics (OpenAI Help Center).

Watch out: this is the floor, not the goal. One prompt a month counts as active.

2. Repeat use

What it is: the share of people with access who used the tool on at least 4 separate days in the last 28.

Where it comes from: the "active days" column. Microsoft's per-user Copilot table counts active days for Copilot Chat, and Google's user-level Gemini report shows active days over the past 28 days. Export, count, divide. If your tool shows no active days, ask in the pulse survey: "On how many days in the past two weeks did you use [tool] for work?"

Why it matters: a habit is the difference between "tried it after training" and "uses it". Four days in four weeks is a modest bar; set your own.

3. Tasks in routine use

What it is: the number of specific tasks that have a written, repeatable method: the task, the approved tool, the prompt or steps, the checklist and an owner.

Where it comes from: a simple list kept by whoever owns AI for the team. A task only counts once someone other than its author has used the method successfully.

Why it matters: this is the closest thing to an adoption number that a manager can act on. "We have six tasks standardized across sales, admin and HR" says more than any usage graph. The AI productivity without losing quality guide explains how a task earns its way onto this list.

4. Time per task against a baseline

What it is: the average minutes a standardized task takes now, including checking, compared with the same task done the old way.

Where it comes from: before training, time five normal examples of the task. Thirty days after, time five more. Rough minutes on a sticky note are fine; invented estimates afterwards are not.

Watch out: don't convert this to dollars or "hours saved per year" for a slide. Report it per task: "client follow-up email, 22 minutes before, 9 minutes now including review." Self-reported time savings without a baseline tend to be optimistic, and people learn quickly what answer the boss wants.

5. Rework rate

What it is: out of a sample of AI-assisted outputs, how many needed a significant correction, either at review or after they were sent.

Where it comes from: once a month, the manager or task owner picks five AI-assisted outputs per standardized task and checks them against the review checklist (facts, figures, dates and names match the source; nothing added that the source doesn't say; tone fits; nothing confidential slipped in). Record "needed a significant fix: yes or no".

Why it matters: speed without this number is how quality quietly drops. If rework is rising while time per task falls, the time savings are partly fake.

6. Reuse of shared prompts and methods

What it is: how many entries in your shared prompt document or library were used in the last month by someone other than the person who wrote them.

Where it comes from: a "used by" column in the shared document, or a one-line question in the pulse survey: "Which shared prompts or methods did you use this month?"

Why it matters: reuse is how one enthusiast's skill becomes the team's skill. A library nobody else uses is a personal notebook.

7. Approved-account use

What it is: the share of people who say they used only approved tools on work accounts for work in the past month, plus the number of AI mistakes reported.

Where it comes from: an anonymous pulse question ("In the past month, did you use any AI tool for work outside [approved tool] on your work account?") and your mistake log. Admin reports can't see personal accounts, so you have to ask.

Read it carefully: a rising number of reported mistakes is often a good sign; it means people trust the reporting process. Zero reports for six months in a team using AI daily usually means people have stopped telling you. The AI security training outline covers how to set up same-day reporting, and the shadow AI guide covers the anonymous survey.

8. Cost per active seat

What it is: what you pay for AI licences per month divided by the number of people who were active (metric 1) or, better, repeat users (metric 2).

Where it comes from: your invoice and the admin report.

Why it matters: it turns "should we renew?" into arithmetic. If you pay for 40 premium seats and 12 people are repeat users, the real price per useful seat is more than three times the list price. Move unused paid seats to the free or included tier before you buy more. For what each licence includes, see our Copilot licence decoder or the ChatGPT Business vs Plus table.

The scorecard (copy this)

Paste it into a spreadsheet, one column per review date. The "healthy" and "red flag" columns are Hero's rules of thumb for a team that has had training and one approved tool for about 90 days. They aren't industry benchmarks; replace them with your own targets after the first review.

#	Metric	How to calculate	Source	Healthy at 90 days (rule of thumb)	Red flag
1	Active-user rate	Active users (28 days) ÷ people with access	Admin report	Most people with access	Falling month on month
2	Repeat use	People active 4+ days in 28 ÷ people with access	Admin report export, or pulse	About half or more of those with access	Active rate high, repeat use low
3	Tasks in routine use	Count of tasks with a written method used by 2+ people	Task list	At least one per team or role	Zero after 60 days
4	Time per task	Average minutes incl. checking ÷ baseline minutes	Five timed examples before and after	Faster on most standardized tasks	No baseline recorded
5	Rework rate	Outputs needing a significant fix ÷ outputs sampled	Monthly sample of five per task	Same or lower than before AI	Rising while time falls
6	Shared-prompt reuse	Entries used by someone other than the author this month	Shared document or pulse	Most entries used by someone	Library growing, reuse flat
7	Approved-account use	People using only approved tools on work accounts ÷ respondents; mistakes reported	Anonymous pulse; mistake log	High, with some mistakes reported	Zero mistakes reported for months
8	Cost per active seat	Monthly licence cost ÷ repeat users	Invoice ÷ metric 2	Close to list price	Several times list price

Add one free-text row at the bottom: "Best example this month." A single concrete story ("the weekly client report now takes one person 40 minutes instead of two hours, and nobody has caught an error in it") is what persuades a sceptical partner or owner.

The four-question monthly pulse

Metrics 2, 6 and 7 need a short anonymous survey when your tool's reports can't supply them. Keep it to four questions so people actually answer:

On how many days in the past two weeks did you use [approved tool] for work? (0, 1–2, 3–5, 6+)
Which shared prompts or methods did you use this month? (List, or "none")
Did you use any AI tool for work outside [approved tool] on your work account? (Yes / No / Not sure)
How confident are you using AI on your own work, 1 to 5?

Ask question 4 before training too, so you have a starting point.

A next stepRather have someone run it? See the public workshop or private team training, or book a free 30-minute call.

The 30/60/90-day review agenda

Put three 45-minute meetings in the calendar on the day training ends. Invite the AI owner, one manager per team and one or two regular users. Each meeting ends with a written decision, not a discussion.

Day 30: is anyone using it?

10 min: metrics 1, 2 and 7. Who has access but hasn't used it? Is that a login, a licence or a "don't know what to use it for" problem?
15 min: each team names its two or three target tasks (if training didn't already) and records the baseline for metric 4.
10 min: what got in the way? Collect the top three blockers from the pulse comments.
10 min: decide one fix per blocker and an owner. Common ones: a 20-minute follow-up for people who missed training, a shared prompt document, or clearer rules on what may go in.

Day 60: is it working on real tasks?

10 min: metrics 1 to 3 against day 30.
20 min: for each target task, time-per-task and rework (metrics 4 and 5). Keep, change or stop each one using the method in the quality guide.
10 min: which methods are ready to share? Add them to the shared library and name an owner for each.
5 min: any mistakes or near-misses? Update the rules if needed.

Day 90: is it worth paying for?

10 min: the full scorecard against days 30 and 60.
15 min: licences. Using metric 8, move unused paid seats down, and add seats only where repeat users are hitting limits.
10 min: pick the next two or three tasks per team for the next quarter.
10 min: decide what happens next: carry on with the monthly scorecard, run a refresher on the weakest area, or widen use to another team. Write the decision down and put the next quarterly review in the calendar.

After day 90, fold the scorecard into the quarterly review in your AI governance framework so adoption, risk and tool approvals are looked at together.

Mistakes that make adoption numbers useless

Counting training attendance as adoption. It measures the calendar, not behaviour.
Celebrating prompt counts. More prompts can mean people are struggling to get a usable answer.
No baseline. Without "before" numbers, every comparison is a guess.
Ranking individuals. People start using the tool to be seen using it. Use per-person data only to offer help.
Measuring only the enthusiasts. If only volunteers answer the pulse, the numbers will look great and be wrong. Ask everyone, anonymously.
Waiting for a dashboard. You don't need analytics software for a 20-to-200-person team. A spreadsheet updated monthly is enough, and you can start this week.

When to bring in help

If the day-30 numbers show low use, the problem is usually that people don't have a clear task or don't feel safe using the tool. Hands-on team training on their own kinds of work tends to fix the first; a short security session fixes the second. If the numbers are fine but nobody can agree what to do next, that's a planning problem: Hero's AI consulting helps set the targets, run the 30/60/90 reviews with you and decide keep, change or stop, usually after an AI Readiness Assessment. The simplest first step for a team is the AI Lunch & Learn, C$899 + HST. You can also run everything on this page yourself.

FAQ: measuring AI adoption

What is a good AI adoption rate?

There's no reliable public benchmark for small and mid-size teams, and vendor figures usually measure their own customers. Set your own target after the first 30 days, and watch repeat use and tasks in routine use rather than the headline active rate.

How do you measure AI training ROI?

Per task, not per company. Take a standardized task, compare time-per-task including checking against the baseline, confirm rework didn't rise, and multiply by how often the task happens. Compare that with seat and training costs. Avoid company-wide "hours saved" figures built on self-reported estimates.

Which AI usage metrics can I get without extra software?

Microsoft 365 admins can use the Copilot usage report, Google Workspace admins have Gemini reports in the Admin console, and ChatGPT Business owners and admins can view workspace analytics. A four-question pulse survey covers the rest.

How often should we review AI adoption?

Monthly for the first three months, with the 30/60/90-day meetings above, then quarterly alongside tool approvals and risk.

How do you increase AI adoption in the workplace?

Name specific tasks per role, give people a shared place for methods that work, make the rules on what may go in clear, and fix logins and licences fast. Low adoption is usually a missing task or a missing permission, not a lack of interest. See how to train employees to use AI for the training side.

Sources checked (October 23, 2026): Statistics Canada, Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026 · Microsoft Learn, Microsoft Copilot usage report and Activity reports in the Microsoft 365 admin center · Google Workspace Admin Help, Review Gemini usage in your organization · OpenAI Help Center, ChatGPT Business workspace roles and access · Ontario, Your guide to the Employment Standards Act: written policy on electronic monitoring of employees.

How to measure AI adoption after training: 8 metrics and a scorecard