Ask an enterprise how its AI rollout is going and the answer arrives as a number of seats. 12,000 licenses deployed. 85% of knowledge workers provisioned. Copilot live across Finance, Legal and Engineering by the end of Q3.
Every one of those statements is true. Each one is auditable against a contract and an identity directory. None of them is an adoption metric, and treating them as one is now the most common error in enterprise AI reporting.
What a license actually records
A license is a purchase order with a name attached. It is generated by a procurement decision and completed by a provisioning action, and both of those events happen in systems that have no visibility into work at all. The contract system knows what was approved. The identity system knows who was granted access. Neither one observes a document, a commit, a ticket, a model, a deal or a deliverable.
This is not a flaw in license data. License data is excellent at its job. It is the most reliable record in the software estate, which is exactly why it gets pressed into service answering questions it was never built for. When a CFO asks whether the AI investment is working and the only complete dataset in the building is the license register, the license register is what ends up on the slide.
AI license utilization is a real metric, and a narrow one
The honest version of a seat count is utilization: of the seats purchased, what share show any recorded activity in the period. That is a legitimate and useful measure. It drives renewal decisions, it catches overprovisioning, and it is the basis of most software asset management practice.
It also has a hard ceiling on what it can support. Utilization can tell you a seat was opened. It cannot tell you what the person did next, whether the output differed from what they would have produced otherwise, or whether the difference was worth the money. Those are questions about work, and the license record contains no work.
The mechanism: why the number rises whether or not anything changes
Here is the part that makes seat counting actively misleading rather than merely incomplete.
Provisioning is a decision made once, centrally, by a small number of people. Behavior change is a decision made continuously, individually, by everyone. The two move on completely different timescales and respond to completely different inputs. An enterprise can provision 12,000 seats in a weekend. It cannot change how 12,000 people work in a weekend, and nothing about the first event constrains the second.
So the adoption curve drawn from license data climbs fast, hits a ceiling near total headcount, and then flattens. It looks like a successful rollout. It would look identical if not one deliverable in the company had been produced differently. The chart has no term in it that could have gone the other way.
The gap is now large enough to be the story
This used to be a theoretical objection. The 2026 software asset management data makes it a measured one.
Flexera's 2026 State of ITAM Report, published June 24 2026 and based on a worldwide survey of 512 professionals performing IT asset management functions, reports that only 31% of organizations have accurate visibility into AI software, against 66% for their SaaS environment generally. The newest and fastest-growing category in the software estate is the one its owners can see least.
The same survey found that only 29% of organizations currently measure the value of AI software, while 59% of respondents report that wasted AI spend increased year over year. Read together, the three figures describe an estate where visibility is thinnest exactly where spend is newest, and where roughly seven in ten organizations have no instrument pointed at return.
One qualification belongs with these numbers, and it cuts against the strongest version of the argument. Flexera's instrument is a self-assessment. The question put to respondents was whether they have accurate visibility into each environment in their IT estate, so 31% is the share of ITAM professionals who believe they can see their AI software, not the share whose coverage was audited and found adequate. Self-assessed capability usually runs optimistic, which means the real coverage is probably worse, but nobody has measured that and this post is not going to assert it.
Flexera's conclusion is stated plainly in the report's own commentary: "License counts alone no longer tell the story. Neither does spend reporting without usage context."
The scale of the older problem, and the limit on borrowing it
It is tempting to reach for the general SaaS waste numbers here, and worth being careful about how far they carry.
Zylo's 2026 SaaS Management Index, released January 29 2026 and built on analysis of more than 40 million SaaS licenses and $75 billion in spend under management, finds that organizations leave an average of 36% of their SaaS licenses unused, measured against industry-recommended utilization levels. That is a large number drawn from a large dataset, and it is about software generally. It is not an AI license figure, and restating it as one would be the same category error this post is arguing against. Several widely circulated "AI licenses go unused" statistics were built by exactly that move, and they do not survive tracing.
What Zylo's data does establish about AI specifically is the entry path. Expense-based SaaS spend rose 267% year over year, and ChatGPT is now the most expensed application in their dataset. AI tools are arriving on employee expense reports rather than through procurement. That is consistent with the low ITAM visibility figure and offers a plausible structural reason for it, though the two findings come from different populations and neither publisher draws the link. You cannot count seats you did not issue.
What each instrument can and cannot answer
| Question | Seat count | License utilization | Controlled cohort comparison |
|---|---|---|---|
| How many people have access? | Yes | Yes | Yes |
| Which seats are dormant and can be reclaimed? | No | Yes | Yes |
| Did the work these people produce change? | No | No | Yes, as an estimate |
| Did it change more than for comparable teams? | No | No | Yes, as an estimate |
| How confident should the board be in that number? | Not applicable | Not applicable | Stated explicitly |
| Is the answer independent of the tool vendor being measured? | Depends on source | Depends on source | Yes, by construction |
The table is not an argument that utilization is bad. It is an argument that the three instruments answer three different questions, and that the first two are being asked the third question because they are the ones available.
What Levos does instead
Levos treats provisioning as context and behavior as the measurement.
The AI Impact signal family captures adoption depth and frequency for every AI tool, output correlation, and power-user breakdown by team and role. Signals are pulled from Microsoft 365, Google Workspace, GitHub, Salesforce and the AI tools themselves, through customer-authorized connections, with no agents installed on employee devices.
Then the part that makes the number defensible. In the sanctioned wording from our own measurement methodology: we compare adopting teams to non-adopting teams, controlling for tenure, role, and tool stack. We disclose the limits of this approach openly. We do not claim full attribution.
That last sentence is not modesty for its own sake. A comparison between adopting and non-adopting teams inside one company is an observational design, not an experiment. Teams that adopt early differ from teams that do not in ways the controls cannot fully absorb, and anyone who tells a CFO otherwise is selling. What the design does beat, clearly and without qualification, is a company-wide before-and-after with no comparison group, which is what most AI ROI reporting currently is.
Every estimate carries a confidence score, and the confidence score is a separate statement from the estimate itself. A thin signal and a low signal are different claims, and a platform that collapses them produces a number that looks identical whether it rests on nine months of dense evidence or three weeks of sparse evidence.
Seat counts stay in the picture. They are the denominator, the renewal lever, and the first thing procurement asks about. Our SaaS Stack Intelligence surfaces utilization, duplicates and abandoned licenses across connected tools for exactly that reason. The argument is not that the number is worthless. It is that the number belongs in the procurement review and not in the board deck labeled "adoption."
If you want the version of this argument that starts from the spend side rather than the license side, we wrote about software stack visibility earlier this year.
The test to run on your own reporting
Take the adoption number currently going to your board. Then ask one question about it: what would have had to happen differently in the business for this number to be lower?
If the honest answer is "finance would have approved fewer seats," the number is measuring procurement. That is a perfectly good thing to measure. It is just not what the slide says it is.
Frequently asked questions
What is AI license utilization? The share of purchased AI seats showing any recorded activity in a period. It is a procurement measure, useful for renewal and rightsizing, and it cannot tell you whether output changed.
Does AI license utilization tell you whether an AI rollout worked? No. It records provisioning and activity. Two organizations with identical utilization rates can have entirely different work underneath, and the license record does not distinguish them.
Why do organizations have less visibility into AI software than into SaaS? Because AI arrived through a different door. Flexera's 2026 State of ITAM Report (n=512, worldwide, June 24 2026) found 31% visibility into AI software against 66% for SaaS, and Zylo reports expense-based SaaS spend up 267% year over year with ChatGPT now the most expensed application.
What share of organizations actually measure the value of AI software? 29%, per Flexera's 2026 State of ITAM Report. In the same survey 59% reported that wasted AI spend increased year over year.
How do you measure AI adoption without a seat count? Observe adoption depth and frequency per tool in the systems where work happens, then compare adopting teams against non-adopting teams controlling for tenure, role and tool stack. Report the estimate with a confidence score and disclose that it is a controlled cohort comparison, not full attribution.
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
Change how work gets measured.
