Why Most Copilot Licenses Go Unused and How to Fix It | Calance

A diagnostic guide to Copilot seat utilization, adoption measurement and license right-sizing.

Assigning a Microsoft 365 Copilot license takes about a minute. Getting a controller to change how she closes the month takes considerably longer. That gap is where Copilot budgets go quiet.

IT marks the rollout complete, finance sees the same line item every month, and somewhere between the two sit a few hundred seats that get opened once or twice and then forgotten. Nobody escalates it, because nothing broke. There is no outage, no failed project and no angry stakeholder, just a recurring cost attached to software nobody is using.

One caution before going further. Several articles claim that 64% of Copilot licenses go unused. That figure comes from reversing a survey statistic which measured something else entirely, and it says nothing about your tenant. Your own usage data is the only source that can tell you how many of your seats are idle. Underuse has at least nine distinct causes, and the remedy differs for each. Some seats need coaching, some need a technical fix, and some should go back into the pool.

Last updated: August 26, 2026

What unused Copilot licenses actually means

Ask three people in the same company how many Copilot licenses are unused and you will get three answers, because each is measuring a different thing. Finance counts seats without recent activity, IT counts accounts that never opened the product, and the business counts people who tried it and gave up. Collapsing every licensed user into a single adoption rate hides the diagnosis entirely, and an organization reporting 45% adoption might have 45% clicking occasionally or 45% who genuinely rebuilt a weekly process. Those two situations call for opposite decisions.

Published research needs the same care. Recon Analytics found 35.8% of paid AI subscribers converted to Copilot against 83.1% for ChatGPT, but that measures platform preference in a survey population skewed toward multi-tool workplaces. Where Copilot is the only platform provided, 68% adopt it. Where three are available, 8% do.

Start with your own tenant data before blaming employees

Most Copilot recovery projects start with a training plan. They should start with an export. Microsoft ships four native reporting sources, and they answer different questions rather than the same question at different resolutions. Build the baseline as a stack and count each of the nine layers separately, because the gap between any two adjacent layers tells you where the problem actually sits.

1. Licenses purchased. Contracted seats taken straight from the agreement, the only number finance already trusts
2. Licenses assigned. Seats attached to a named person in the tenant, which is usually lower than purchased
3. Technically enabled. Eligible base plan, supported update channel and working apps, all three confirmed per user
4. Active users. Any qualifying action inside your chosen window, and the window setting decides this number
5. Returning users. Action recorded in more than one week, the sharpest early signal that first use took hold
6. Habitual users. Consistent use across most weeks in the measurement period, at any volume
7. Power users. Frequency and consistency together, not one or the other in isolation
8. Workflow adoption. A named recurring process that demonstrably changed, described in one sentence
9. Business outcomes. An operational or financial number that moved, baselined before the pilot began

Two details change everything. The Viva Insights report lets you customize the active-user definition, so whoever picks that setting quietly decides your adoption rate. The Copilot readiness report arrives within 72 hours and carries up to 72 hours of latency, so treat it as last week. Exports also anonymize user and group details by default. Calance handles this groundwork as part of

The first failure often happens when licenses are assigned

Plenty of Copilot seats were doomed on the day they were handed out, and no amount of later training recovers them. Five assignment models cause most of the damage.

A better filter checks six things before a paid seat is attached: whether the person lives in Teams, Outlook and Office documents, whether they repeat a task weekly, whether you can name it in one sentence, whether their content is in reasonable shape, whether their manager will reinforce it, and whether they will change how the work gets done. Our Copilot readiness guide walks through the assessment. Copilot Chat also comes free with eligible subscriptions, so you can watch someone use enterprise AI before buying a seat.

Decide between Copilot Chat and a paid seat before you buy

Copilot Chat comes at no extra cost with eligible Microsoft 365 subscriptions, which gives you a way to watch someone use enterprise AI before buying them a paid seat. The progression from Chat to demonstrated use cases to a paid license gives you evidence before spend, and gives the employee a reason to expect the upgrade rather than receive it unannounced.

That caveat matters when you qualify people. Chat coverage is narrower for in-app work, so anyone whose recurring task lives inside Word or Excel cannot be fully assessed through Chat alone, and treating a quiet Chat user as disqualified will lose you genuine candidates. Use the tier to filter out people who never form a habit at all, not to prove that the people who do need nothing further.

Copilot has nothing to attach to when the workflow was never defined

Workflow fit is the cause training cannot reach, and it is probably the most common one on this list. A person can be fully licensed, fully trained and perfectly willing, and still have nothing in their week that Copilot improves. Feature-led rollouts open with capability, telling people Copilot can summarize, draft and analyze. Workflow-led rollouts open with a job, naming how the Thursday status update gets prepared today. Both descriptions are accurate, and only one tells an accounts payable clerk what to do differently on Tuesday. Run every candidate workflow through six questions.

Workflow qualification

• Recurring. It happens weekly or monthly, so the person never has to relearn the approach from scratch. Weekly tasks build habits and daily tasks build them faster.
• Information-heavy. It involves reading, gathering or synthesizing existing material rather than originating something from nothing.
• Time-consuming. There is enough time inside it to be worth reclaiming, measured in tens of minutes rather than two.
• Grounded in Microsoft 365. The source material lives in Teams, Outlook, SharePoint or OneDrive, where Copilot can actually reach it.
• Suitable for AI assistance. A draft or a summary genuinely moves the task forward, rather than creating something that needs rewriting anyway.
• Reviewable. A person can check the output before it goes anywhere, and knows what checking it actually involves.

Meeting follow-up, weekly reporting, proposal drafting and document summarization usually qualify. Test that against a department calendar rather than assuming it, since meeting follow-up matters enormously to a project manager and barely registers for a field engineer. Where a task is structurally broken, workflow automation beats layering AI on top of it.

Training fails when it teaches Copilot instead of the job

Microsoft hit this inside its own Sales and Service organization across 60,000 employees. Broad guidance scaled easily and connected poorly. What worked was role-based immersion: prompts grounded in real workflows, tied to specific responsibilities.

Notice what every row includes. The human check belongs inside the training itself, never in a footnote delivered afterward, and teams that skip it tend to manufacture their own trust problem within weeks. A few structural choices make these sessions land. Keep groups under about 15 people so everyone works on their own material, run the session inside the app where the work actually happens, and hand out three or four prompts tied to that week's deliverable instead of a reference library nobody opens. Schedule follow-up office hours two weeks later, when the first real questions have surfaced, and ask the manager to attend, because their presence signals whether any of this is optional. Plan a refresher each quarter and keep prompt sets under version control, since Copilot ships new features often enough that a March session describes a slightly different product by September.

Trust, permissions and information quality can stop adoption

Some employees stopped using Copilot because they tried it and did not believe the answer, which is a different problem from never having opened it. Recon Analytics found that among users who tried Copilot and stopped, 44.2% cited distrust of answers, against 42.8% for Gemini and 40.6% for ChatGPT. Those are perception measures inside a survey population rather than an audit of output quality, which makes them weaker evidence and still the wrong numbers to wave away. Six signals are worth counting before you expand seats any further.

Content estate signals

• Broad sharing through Anyone links, Everyone groups and organization-wide access grants that nobody has revisited
• Large audiences carrying more permissions than the work actually requires, usually inherited from a long-closed project
• Broken permission inheritance and improvised access models layered on top of each other over several years
• Sensitive content sitting with weak or entirely missing sensitivity protection applied
• Unlabeled or publicly scoped sites that no owner has reviewed since creation
• Ownerless or inactive sites still holding live content that Copilot will happily surface

Part of the perception problem is grounding. An employee who asks about the expense policy and receives the 2022 version concludes the tool is unreliable. The tool was reliable and the content estate was not. A pilot of 50 barely stresses the estate, while 5,000 users will find whatever it holds. Calance approaches remediation through SharePoint consulting, because cleanup is content architecture work before it is an AI project.

Teach users which tasks Copilot should and should not be trusted with

Expectation setting handles most of what governance work cannot. Users need to know which bucket a task falls into before they start, because output confidence rises when expectations match reality. Three tiers cover almost everything a knowledge worker will bring to it.

• Assistance tasks. Drafting, summarization and meeting follow-up. An error here costs a quick edit and nothing more. Copilot output is a starting point the person shapes, and the review is fast enough that it never feels like overhead.
• Verification tasks. Analysis, customer-facing output, and anything financial, legal or regulated. Every figure gets checked against the source system before it moves. The time saved is in assembly, never in the checking, and teams that skip the check manufacture their own trust problem.
• Judgment tasks. Final high-stakes calls stay with a person. Copilot assembles the inputs and the human owns the decision. Nothing produced here leaves the building without a named person having read it end to end.

Teach that split during enablement rather than burying it in a policy document nobody opens, and repeat it in every refresher session. Output confidence rises when expectations start matching reality, and the failure mode where someone forwards an unchecked summary to a client becomes far less likely. That protects the program politically as much as it protects the client, which matters when the first renewal conversation arrives.

Separate technical blockers from adoption resistance

Before anyone schedules more change management, check whether the user could actually use the product. Underuse arrives in three categories that look completely identical in a usage report and need entirely different responses. The readiness report separates the technical category cleanly, since its user table shows per person whether a license is assigned, whether the device sits on an eligible update channel, and whether they are active in the apps where Copilot adds value.

Work them in order for a practical reason. Technical fixes are cheap, fast and permanent. Governance work is slower and repairs problems that predate Copilot. Behavioral change costs most and shows results slowest, so it deserves a population already cleared of the other two. Calance treats infrastructure and endpoint operations as part of a recovery plan rather than a separate workstream.

What sustained Copilot usage actually looks like

Curiosity arrives free. Microsoft found widespread initial experimentation and much thinner sustained usage. The 2026 Work Trend Index, surveying 20,000 workers across 10 markets, reports organizational factors carrying roughly twice the AI impact of individual ones, 67% against 32%. Microsoft frames that as an association rather than a causal measurement.

The champion role

• Tests workflows in their own job before recommending them to anyone else on the team
• Documents the examples that worked, including the prompt and the surrounding context that made it work
• Surfaces the failures too, so peers know in advance where the tool struggles and stop expecting magic
• Answers questions from the people sitting near them, in the moment the question actually arises
• Shares validated prompts tied to one specific recurring task rather than a general reference library
• Feeds recurring problems back to whoever owns the program, so fixes happen once instead of locally
• Holds office hours that appear on a calendar and survive a busy week

Give the role about two hours a week, name it explicitly in the person's objectives, and rotate it annually so it never becomes a permanent tax on one willing volunteer. Pick for credibility rather than enthusiasm. The most effective champions tend to be mid-level people who already field software questions from their colleagues and usually sit outside IT entirely. Their advantage is proximity, because they know what the Thursday report actually involves, so their examples survive contact with reality. Give them a dedicated channel, early access to new capability, and a route to escalate recurring problems to whoever owns the program. Recognition matters more than budget here, and champions asked to cheerlead on top of a full job quietly stop by the second month.

Measuring habit instead of first use

A user who opened Copilot once last month and a user who runs it three times a day both count as active under most default settings. Reporting that treats them identically will send you to the wrong intervention, which is why the Copilot adoption report in Viva Insights carries frequency and consistency together rather than picking one of them.

The rolling window is configurable and the power user threshold adjustable, so record whichever you choose. App concentration deserves a note: in the UK cross-government trial, Teams use peaked at 71% while Excel and PowerPoint topped out at 23% and 24%. A scorecard that averages across apps hides that. Building this into a recurring view is standard analytics and dashboard work.

Where Copilot ROI actually begins

Usage and value are separate measurements, and the distance between them is where most Copilot business cases quietly fall apart. A license purchase creates cost on day one, while value appears somewhere much further along a chain, and most reporting jumps straight from the first link to the last without checking the ones in between. Every link is a place the chain can break. Usage without workflow change produces activity and nothing else, while workflow change without an operational outcome produces spare minutes that get absorbed into the day.

1. License spend. Cost lands immediately on assignment and does not vary with how much the person uses the product
2. Usage. Activity gets recorded, which on its own proves nothing beyond the fact that someone opened something
3. Workflow change. A named recurring process is demonstrably done differently, and the person can describe the difference
4. Time or capacity effect. Minutes or hours are actually released from that process rather than absorbed elsewhere in the day
5. Operational outcome. Cycle time, backlog age, throughput, response time or rework rate moves against its baseline
6. Financial effect. The moved operational number reaches a budget line somebody owns and can point to

Evidence on the last link is thinner than the marketing suggests. The UK cross-government Copilot experiment across roughly 20,000 users reported an average self-reported saving of 26 minutes per day, with 17% reporting no clear saving. Separately, a field experiment across 66 firms found engaged users spending two fewer hours weekly on email with no detected shift in output. Baseline an operational measure before the pilot starts, as with any

When reclaiming a Copilot seat is the right call

The Microsoft 365 Copilot enterprise add-on is listed at $30 per user per month on annual commitment, checked in August 2026. A seat costs the same whether the person uses Copilot daily or opened it once in March. Multiply your own idle count by your own rate, then apply a decision to each seat.

Announce the reassignment policy at rollout rather than at the first reclaim, since a seat people know will be reviewed gets treated very differently from one they assume is theirs forever. Record the reasoning against each seat so the next review starts from evidence rather than from scratch, and so the person making the call in six months is not the only one who understands why the last decision went the way it did.

A practical operating model for fixing Copilot underuse

Everything above assembles into a loop that runs on a schedule. It has no finish line. Each stage answers one question and hands a specific output to the next, which keeps the work from collapsing back into a training exercise.

1. Measure. Pull readiness, usage and adoption data. Fix the active-user definition and write it down before anyone forms an opinion about the result
2. Segment. Sort licensed users into non-users, novice, habitual and power. Break the view down by department and by app, because organization-wide averages hide everything useful
3. Diagnose. Run every low-use seat through the technical, information and adoption split. Pair telemetry with a short survey, since usage data alone leaves blind spots about where people got stuck
4. Prioritize. Pick the roles and workflows worth supporting. Depth in three functions beats a thin layer across twelve, and the thin layer is what most programs default to
5. Enable. Role-based sessions on real documents, a short prompt set per function, named champions with time allocated, and explicit review rules taught alongside the prompts
6. Govern. Repair permissions, retire stale content, apply sensitivity labels, and tell users plainly what Copilot can and cannot reach on their behalf
7. Measure value. Track returning and habitual users first, then the named workflows, then the operational measure you baselined before any of this started
8. Right-size licenses. Apply the seat decision matrix, record the reasoning against each seat, then return to stage one with a cleaner population

Run it quarterly for the first year and semiannually afterward. Assign one owner across all eight stages, because the common failure is IT owning measurement while the business owns enablement, with nobody owning whether a seat should exist. Scale governance to the organization: a 200-person company needs a named owner and a monthly data review, not a formal center of excellence. The loop extends past Copilot, since agentic AI reads from the same content estate and carries the same measurement problem.

Microsoft 365 Copilot adoption FAQs

01 Why do employees stop using Microsoft 365 Copilot after getting a license?

Usually because no recurring task was identified for them to change, or because an early answer felt unreliable enough that they went back to the familiar method. Technical blockers such as an unsupported update channel also stop people quietly, without generating a complaint. Each cause needs a different remedy, so diagnose the population before adding more training to it.

02 How can we identify unused Copilot licenses?

Use the Copilot readiness and usage reports in the Microsoft 365 admin center, which cover license assignment, eligibility and app activity over 28 days and export to CSV for analysis. The Viva Insights adoption report adds usage-level segmentation across a rolling multi-week window. Check whether your export is anonymized before planning any conversation about specific teams.

03 What is a good Microsoft 365 Copilot adoption rate?

No single percentage applies, because the number changes entirely with how you define an active user, and that definition is a configurable setting rather than a fixed standard. Measure sustained usage instead: the share of licensed users returning weekly, the share reaching habitual status, and the share running at least one named workflow that produced a measurable result.

04 What is the difference between Copilot Chat and Microsoft 365 Copilot?

Copilot Chat comes at no additional cost with eligible Microsoft 365 subscriptions and provides enterprise-protected AI chat and agents. The paid license adds work-grounded assistance inside Teams, Outlook, Word, Excel and PowerPoint. Chat coverage is narrower for in-app work, so verify current entitlements against Microsoft documentation before planning a qualification path around it.

05 Should every Microsoft 365 employee receive a paid Copilot license?

Rarely. Paid seats pay off where someone has recurring, information-heavy work grounded in Microsoft 365 and a manager willing to reinforce the change in team routines. Employees without that profile often get sufficient value from Copilot Chat, and can be reassessed later against actual usage evidence rather than an assumption made at rollout.

06 How long should we give an employee before reclaiming an unused license?

Set an internal review period instead of borrowing someone else's threshold from an article. A reasonable one starts only after the person has had a defined use case, role-based enablement, and at least one full measurement cycle behind them. Many organizations review high-cost add-on seats more frequently than standard licenses, which is a defensible position to take.

07 Can Copilot licenses be reassigned?

Yes. A license can be removed from one user and assigned to another, which turns seats into a rotating asset rather than a permanent grant nobody revisits. Announce that policy at rollout rather than at the first reclaim, because a seat people know will be reviewed tends to get treated very differently from one they assume is theirs indefinitely.

08 Does Copilot adoption depend mainly on training?

Training addresses one cause among several, and usually not the largest one. Workflow fit, technical readiness, content and permission quality, output trust, leadership modeling and measurement discipline all affect whether a seat gets used. A rollout that fixes only training tends to plateau within a quarter, then gets described internally as an adoption problem when it was never one.

09 How should businesses measure Copilot ROI?

Separate three things: usage, workflow change and business effect. Baseline an operational measure such as cycle time, backlog age or response time before the pilot starts, then track whether it moves against that baseline. Treat self-reported time savings as directional evidence only, since recovered minutes become value only when something specific absorbs them.

10 Where should an organization start if Copilot adoption has already stalled?

Start with tenant data instead of a training plan. Establish how many seats are genuinely inactive under a definition you have written down, separate technical and governance blockers from behavioral ones, then fix the cheapest categories first. Enablement works far better on a population that already has working access and trustworthy content behind it.

11 Will Copilot surface documents employees were never meant to see?

Copilot answers from content it reaches through Microsoft Graph, scoped to the permissions each user already holds. It does not grant new access, but it does make existing oversharing searchable in seconds. A decade of broad sharing links and broken inheritance becomes visible at once, so audit permissions before scaling seats rather than after.

12 Who should own Copilot adoption internally?

One named person across measurement, enablement and licensing. The common failure is IT owning the usage reports while the business owns training, with nobody owning the decision about whether a seat should exist at all. A 200-person company needs an owner and a monthly data review, not a formal center of excellence.