Unused AI Licenses: How to Find Unused Seats

An unused seat is an AI license that is paid for and assigned to a person who does not use it, or uses it so rarely that the cost is not justified. Finding it requires an operational definition of "inactive", usage data cross-referenced with the list of assignments and a check for false positives before any revocation. Reclaiming is the last step and must be handled with advance notice.

How to define an inactive seat: thresholds and time windows

Without a shared definition, every department has its own idea of inactivity and the discussion drifts to individual cases. The definition has two parameters: the usage threshold and the observation window.

• Usage threshold. The minimum that makes a seat "active": at least one work session, a number of interactions, a certain number of access days in the period. The lower the threshold, the more cautious the analysis.
• Observation window. The period over which you measure. Windows of 30 days capture too much noise (holidays, work peaks), while beyond 90 you risk holding idle licenses for a long time.

A frequent starting point is to define three classes:

1. Inactive: no use in the last 60 days.
2. Sporadic: use below the threshold expected for the role, for example less than one access day per week.
3. Active: regular use, at or above the threshold.

These are example values, to be calibrated to the company: the right number is the one that cleanly separates people who use the tool in their work from those who opened it once. A threshold must apply to all comparable tools, otherwise spend comparisons lose meaning.

Where to find the data: vendor consoles and access logs

You need two lists to cross-reference: who has a license and who uses it.

List of assignments. It comes from vendor admin consoles, from the corporate identity system (the groups or roles that activate the license) and from IT spreadsheets. Often the three lists do not match, and the difference is already a result: licenses paid for with no assignee, or assigned to people who have left the company.

Usage data. Vendor admin consoles expose information on user activity to varying degrees: depth, refresh frequency and retention period change from tool to tool. Where usage activity is not available, use a proxy: identity system logs show when a person authenticates to the tool, even if not what they do with it.

A few precautions:

• check for each tool how long data is retained, to avoid discovering that the 60-day window cannot be reconstructed;
• normalize identifiers (the same person can appear under different addresses);
• aggregate by department and role before sharing the analysis.

If data is collected at individual level, the processing must be assessed with the DPO and union relations: the article on monitoring AI use and the Italian Workers' Statute covers it.

False positives: when an inactive seat is not waste

A seat that looks inactive should not be revoked automatically. The most common false positives:

• Seasonal use. Functions such as year-end accounts, planning or tenders have peaks concentrated in certain months. A 60-day window in the low period classifies them as inactive.
• Specific roles. People who use the tool for rare but high-value tasks (a legal analyst, an audit role) may have few accesses and a lot of benefit.
• Long absences. Parental leave, illness, sabbaticals and long assignments produce inactivity without any lack of interest.
• New hires and recent licenses. A license assigned a few weeks ago has no meaningful history yet.
• Use on untracked channels. If part of the use goes through an integration or a channel the console does not record, activity is underestimated.

The remedy is a verification step with the department head before revocation, and a clear exclusion rule for the categories above. A 90-day threshold for seasonal roles is often more useful than a long list of exceptions.

Worked example: what reclaiming is worth

The example is hypothetical: the values are invented for illustration.

A company has 800 seats at €240 a year each, for a spend of 800 × €240 = €192,000. With a 60-day threshold, 96 seats turn out to be inactive, which is 12% (96 ÷ 800 = 0.12). Gross value: 96 × €240 = €23,040.

The check with department heads reveals 16 false positives (long absences and one seasonal department). The recoverable seats are 96 − 16 = 80, for annual savings of 80 × €240 = €19,200, about 10% of spend (19,200 ÷ 192,000 = 0.10).

The step from €23,040 to €19,200 shows why you should not present the gross figure as savings: the €3,840 difference (16 × €240) is the share the raw analysis would have taken away from people entitled to keep it. You still need to verify whether the contract allows reducing seats before renewal: the renewal negotiation checklist covers this.

The reclaim process, without irritating people

Removing a tool without explanation generates resistance and, worse, pushes people toward personal accounts outside control (see shadow AI: risks and how to manage it). A process that works:

1. Communicate the criterion in advance. A message to everyone explains that licenses are limited resources, defines threshold and window and announces the periodic review.
2. Produce the candidate list. Apply the threshold to the data and remove known false positives.
3. Verify with department heads. Each head confirms or contests the list for their department by a deadline.
4. Notify the people concerned. An individual message with advance notice (for example 14 days), the reason and how to keep the license.
5. Offer an activation attempt to sporadic users. Short training or a targeted use case before removal.
6. Revoke and document. Record the date and reason, and free the seat for someone on the waiting list.
7. Guarantee restoration. Whoever needs it must be able to get it back within a few days, without bureaucracy.

Fast restoration is what makes revocation acceptable: it lowers the cost of a mistake and removes people's reason to bypass the process. It also helps to name a contact person and a maximum response time, so requests do not get lost between IT and the department.

How to make the check periodic

A one-off analysis recovers licenses, but drift starts again immediately: new assignments, role changes and departures bring back inactive seats within a few months. Set a quarterly review with the same threshold, and a deeper check before each renewal. Activation rate and cost per active user, described in the article on the real cost per AI seat, are the indicators to follow over time.

Where Delphis fits

Delphis is a vendor-independent platform that connects spend, adoption, impact and regulatory obligations. It cross-references purchased licenses with actual usage across multiple tools, flags inactive and sporadic seats, simulates the savings of different scenarios and measures impact against a baseline. Data on people stays aggregated by role or department. To see how the reconciliation of costs and usage is set up, see the Spend page.