Search for AI usage analytics tools and the results split in two, and neither half is yours. Some are business intelligence products that happen to run on AI. The rest are employee-monitoring platforms that count minutes inside an application window and call it adoption.
For a company of 50 to 500 people the best option is usually the admin console that came free with the assistant seats you already bought. The paid layers earn their price in two specific situations, and this is how to tell whether you are in one of them.
Four layers, four different kinds of evidence
Every product sold under this heading sits in one of four layers, and the layer decides what evidence you hold when your CFO asks whether the AI enablement spend worked.
| Layer | Where the data comes from | Typical cost | What it proves | Where it fails |
|---|---|---|---|---|
| Native admin console | The vendor’s own logs | Included with your seats | Who has the habit, tool by tool | One tool at a time, no quality signal |
| Cross-tool analytics | Connectors into each console, plus collaboration data | About $2,500 a month at 200 people | One view across tools, cut by team | Still usage, in a single pane |
| AI discovery | Email metadata, OAuth grants, network or browser telemetry | From about $750 a month | Which AI tools exist, personal accounts included | Nothing about skill |
| Sign-in logs plus one graded work sample | Your own logs and a manager’s hour | Free | Whether the work changed | Needs a human to grade it |
Read the fourth column. Three layers measure the same thing at rising resolution and cost, and only the last produces evidence about the job. That gap between recorded activity and demonstrated capability is the argument in our AI enablement measurement playbook. It also explains McKinsey’s 2026 result: across 1,719 respondents, 80% said AI had improved their own productivity while 6% credited it with 5% or more of company EBIT. Activity data measures the first number and never reaches the second.
Start with the console you already pay for
Every major assistant publishes adoption data to admins without a separate invoice line, and the detail is better than buyers assume.
Microsoft 365 Copilot. The admin center usage report covers enabled users, active users, active users rate, active agent users, total prompts, and average prompts per user, broken out by app, over 7, 30, 90, or 180 days. A day’s data is typically available within 72 hours of that day’s end in UTC. The Copilot Dashboard in Viva Insights adds readiness, adoption, impact, and sentiment pages.
Claude. Enterprise organizations get an analytics dashboard plus an analytics API covering per-user daily activity, active-user summaries, seat counts, and project, skill, and connector adoption. Anthropic’s July 2026 release added usage and cost by group and by individual, filtered by the SCIM groups IT already manages, with spend alerts at 75 and 90 percent of an organization limit.
Google Workspace. Gemini reports live in the Admin console under generative AI, at organization and user level, covering per-feature active users and per-app interactions. A February 2026 release added reporting on users who have hit their usage limits, which doubles as a signal that someone needs a larger plan.
ChatGPT. OpenAI ships workspace analytics in the admin console and an admin endpoint for daily usage rows. Confirm the metric list and plan eligibility with your account contact, since the administrator documentation is not publicly indexable.
Four consoles, four schemas, no new line item. For a company running one or two assistants, this layer is the product.
The license cliff nobody warns you about
Two gates decide whether the free layer is complete for you, and both catch mid-market buyers. The Copilot Dashboard in Viva Insights needs no paid Viva Insights license and no Copilot license to open, but what it shows depends on how many Copilot licenses the tenant holds. A tenant with 1 to 49 assigned licenses gets the readiness page plus tenant-level and group-level adoption and impact metrics. It does not get agent insights, internal benchmarks, intelligent summaries, survey-based sentiment, week-level or month-level trendlines, or the group view and delegation that let a manager see their own team. At 50 assigned Copilot licenses, or 50 Viva Insights licenses, all of it unlocks.
That lands badly on a common pattern: a 220-person company buys 30 seats for a finance pilot, runs it a quarter, then cannot produce a trendline or a team comparison. The data is not missing. The reporting tier is.
The second gate is plan tier. Claude’s engagement and adoption API is available to Enterprise organizations, not to Team or Pro, so a 150-person company on a team plan cannot pull that data into a report of its own.
Check both before writing a measurement plan. Adding ten seats to clear a threshold is cheaper than buying a platform to work around one, and the tier decides whether you can compare yourself to published adoption benchmarks at six and twelve months.
Why the impact number cannot disagree with the usage number
This is the most useful fact in the category, and no comparison page states it.
Microsoft’s Copilot Dashboard reports an impact figure called Copilot assisted hours, computed from usage counts multiplied by fixed assistance factors. Creation actions are credited at 6 minutes each, from a study of 147 knowledge workers who drafted a blog post 6 minutes faster with Copilot. Search and summary actions also earn 6 minutes, from a study of 163 workers retrieving information. Meeting summaries count the duration of the meeting summarized. Copilot assisted value is that hour total times an hourly rate, defaulting to $72 from US Bureau of Labor Statistics employer-cost data.
Microsoft documents all of this openly and calls the result a broad approximation. The company is being candid; the deck that lifts the dollar figure onto a board slide is not.
The consequence is structural. Because value is a fixed multiple of actions, any program that increases clicking increases reported value, whether or not one deliverable improved. Put it in front of finance as a result and the first person to read the methodology page takes the program down with it.
Use it for what it is: an activity index, fine for comparing this month to last. Evidence that an output changed has to come from the output, and our walkthrough of a finance-ready ROI model sets out that arithmetic.
When a cross-tool layer earns its price
Paid aggregation exists because four schemas cannot be compared, and because no console knows that the person who stopped using an assistant also changed managers.
Worklytics is the clearest example and the most focused on adoption. Free up to 100 users, then $2,500 a month covering up to 200 users and $10 per user per month above that, joining AI usage with calendar, email, and project signals so the unit of analysis is a team, not a tool.
ActivTrak arrives from workforce analytics. List pricing runs $10, $15, $17, and $21 per worker per month across four annual tiers, with an AI Insights add-on at $4 per worker per month.
Do the arithmetic first. At 200 employees the $17 tier plus the add-on is $4,200 a month, about $50,000 a year, which exceeds the market range for a full hands-on LLM fluency program: counting money rather than teaching money. The wider version of that trade runs through our guide to diagnosing a program that did not work.
Two conditions make aggregation a sound buy: three or more assistants in parallel, and a named person who reports the number on a fixed cadence. Without the second you have bought a dashboard nobody opens.
One thing pricing pages do not raise: several products here are marketed as employee monitoring, with per-person activity timelines as the headline feature. Announcing one in the quarter you ask people to experiment openly is a trade, not an addition.
What no dashboard can see
Every console here reports on sanctioned accounts only. Netskope’s Cloud and Threat Report for 2026 found 47 percent of workplace generative AI users still working through personal accounts, down sharply from 78 percent a year earlier. Close to half the use you are trying to measure may sit outside the console you are reading.
Discovery tools look somewhere other than the vendor’s logs. Nudge Security detects AI tools from email metadata and OAuth grants, which surfaces anything tied to a corporate identity whatever the device, at $750 a month for up to 150 users and $5 per user per month from 150 to 1,500. SaaS management platforms such as Torii, Zylo, and Productiv get there through expense feeds and single sign-on logs. Network and browser layers such as Netskope, Cloudflare, and LayerX see the traffic directly, usually on a security budget.
Buy this layer when the question is risk: unreviewed tools touching customer data, or a policy you cannot enforce because you cannot see the breach. One anonymous survey question a quarter tells you what share of the team uses personal accounts, for nothing.
Three constraints that decide whether a tool is usable
Latency. Microsoft’s report lags up to 72 hours, Claude’s engagement data runs roughly a day behind, and Claude’s cost figures keep reconciling for about seven days after a month closes. Weekly reporting works; a daily adoption standup does not, and month-end numbers quoted on the first will move.
Privacy defaults. Microsoft 365 usage reports conceal user, group, and site names by default, only a Global Administrator can turn concealment off, and the change is written to the audit log. That inverts the usual framing: these products ship private, and someone decides to make them identifiable. Say which way you set it before anyone asks.
Minimum group size. Group metrics are suppressed below a configured minimum, so a six-person finance team can show a dash where a number should be. Microsoft’s satisfaction rate needs 30 responses from five unique users in a rolling 28 days before it displays at all. Plan department reporting around the smallest department you care about.
Microsoft also publishes thresholds worth borrowing: a power user averages 15 or more Copilot actions a week across at least 9 of the past 12 weeks, and a habitual user is 1 to 14 on the same bar. Both beat the bars companies invent for themselves.
What to do, in order
- Open the console you own. Record seats assigned against seats active, weekly active users among the people you trained, and the share clearing a sustained-use bar. Use Microsoft’s definitions so the bar is not yours to defend.
- Ask one anonymous question. What share of the team uses a personal AI account for work? That tells you how much of the console is the whole picture.
- Timestamp one output. Pick a recurring deliverable and record how long it takes and how good it is. This is the only measurement here that can contradict the dashboard, and the method matches reading a pilot result before you scale it.
- Only then price a paid layer. Aggregation if three or more assistants are in play and someone owns the weekly number. Discovery if the question is risk. Neither, if what you are missing is a graded work sample.
Those steps take a week and resolve most of the question.
Frequently asked questions
What are the best AI usage analytics tools for a mid-market company?
The native admin consoles, for most companies of 50 to 500 people, because they hold first-party data and cost nothing: the Microsoft 365 Copilot report, the Claude analytics dashboard on Enterprise plans, Gemini reports in the Google Workspace Admin console, and ChatGPT workspace analytics. Add Worklytics or ActivTrak once three or more assistants are in play.
How much does AI usage analytics software cost?
Native consoles are included with your seats. Cross-tool analytics starts near $2,500 a month for 200 users on Worklytics, or $21 per worker per month for ActivTrak’s $17 tier plus its $4 AI Insights add-on, about $50,000 a year at 200 people. Discovery starts near $750 a month. Prices verified October 2026.
Can I measure AI adoption without buying anything?
Yes, and for most companies that is the right first move. Your console reports activation, active users, and prompt volume, your sign-in logs show which AI services people authenticate to, and one anonymous survey question covers personal-account use. The only missing piece is a graded work sample, which costs a manager an hour.
Does the Copilot Dashboard work if we only bought a few licenses?
Partly. It opens for any Microsoft 365 business or enterprise tenant with an active Exchange Online account, with no paid Viva Insights or Copilot license needed. But a tenant with 1 to 49 assigned Copilot licenses loses benchmarks, agent insights, sentiment, trendlines, and the group-level manager view. All of that unlocks at 50 licenses.
Is Copilot assisted value a real ROI number?
No. Microsoft computes it as usage actions multiplied by fixed assistance factors, 6 minutes per creation or search action, multiplied by an hourly rate defaulting to $72 from Bureau of Labor Statistics employer-cost data. It is activity expressed in dollars and it rises whenever clicking rises. Treat it as a month-over-month index, never as evidence for finance.
How do I track AI tools employees use on personal accounts?
Discovery tools rather than usage dashboards: Nudge Security reads email metadata and OAuth grants, SaaS management platforms use expense and single sign-on data, and network or browser layers see the traffic. The cheap version is an anonymous quarterly survey question, enough when the question is enablement rather than risk.
Will usage analytics tell me whether our AI enablement worked?
No. Usage tells you who formed a habit, which is necessary and not sufficient. Capability needs a separate instrument, which is why the shortlist of AI fluency assessment options is a different purchase, and why a dated before-and-after on one real deliverable beats both.
Do we need to tell employees we are measuring their AI usage?
Treat it as required whatever your jurisdiction says. Microsoft 365 reports conceal names by default and unconcealing them is a logged administrative act, a sensible model to copy: report at team level, say so in writing before the first report runs, and name who can see individual rows.
What is the single metric worth reporting monthly?
The share of the people you trained who used an assistant on real work last week, paired with one dated output comparison per quarter. The first is free from the console you own, the second costs a manager an hour, and together they answer what the paid stack only approximates. That pairing is the core of our case for building capability rather than buying licenses.
Key takeaways
- Four layers exist and three measure the same thing. Native consoles are free, aggregation runs about $2,500 a month at 200 people, discovery starts near $750, and a graded work sample is free and the only layer that proves the work changed.
- Open the console you already pay for first: Microsoft, Anthropic, Google, and OpenAI all publish admin adoption data at no extra charge.
- Check the thresholds first: under 50 Copilot licenses the Viva Insights dashboard loses benchmarks, trendlines, sentiment, and manager views, and Claude’s analytics API is Enterprise-only.
- Copilot assisted value is mostly actions times 6 minutes times a default $72 rate, so it cannot contradict a usage increase. It is an activity index, not a result.
- Close to half of workplace AI use still runs through personal accounts, so every console here reports on a partial population.
- Buy aggregation only when three or more assistants are in play and a named person reports the number weekly.
To have this run against your own consoles, book a free AI readiness call. We will say which of your dashboards already answers the question, which threshold you are under, and what the missing measurement costs.
Dirk Jan van Veen, PhD
