On November 18, CloudEagle.ai released a "State of AI Governance 2026." The sample was more than 600 CIOs and CISOs, covering the United States, the United Kingdom, Europe, and Latin America. There was one sentence in the report that I read twice:
After an employee leaves, their AI access permissions remain active for an average of another 60 days before being terminated
In the same report, there is another scenario. When conducting an AI asset inventory for a client, the number reported by the system made the CIO think something had gone wrong—he remembered approving only 12. CloudEagle founder Nidhi Jain's exact words were: “It was never hidden; it was just that no one had ever measured it.”
These two things are actually one thing: we manage the entry and exit of 'people' in great detail, but we have not managed the entry and exit of 'identity.' And AI is precisely connected into enterprises in the form of identity.
1. First, lay out the numbers
This report is a vendor-commissioned survey; CloudEagle itself sells AI governance and SaaS management, so the figures should be treated as a directional reference. But when several sets of numbers are put together, the implications are quite clear.
| Indicator | Numeric value | pronunciation |
|---|---|---|
| How many AI tools do large enterprises run on average | 152 (89 a year ago) | It rose by 70% in a year, and the ledger basically can't keep up. |
| Which systems can the agent access? | 61% customer data, 58% code repositories, 54% HR systems | What most agents receive is not “read-only scraps” |
| Organizations with a formal AI procurement process | 42% yes, 58% no | Business buys and deploys it themselves, and IT only finds out afterward. |
| Organizations with a dedicated AI compliance officer | 31% yes, 69% no | When something goes wrong, you can't find anyone to sign. |
| “Well prepared” for regulations already in effect | 13% | The EU AI Act, South Korea's AI Act, and California SB 53 have all taken effect |
| Organizations that have performed risk classification for AI systems | 43% have done it, 57% have not | Classification is the first step in all regulatory requirements. |
| Per capita AI spending (large enterprises) | $2,150 per year | about five times that of small businesses, yet the account is often booked under the business department |
What concerns me most is not the number 152, but the gap implied by "only 12 approved." It shows that the path by which AI enters enterprises is completely different from the path by which software entered enterprises back then: software had to go through procurement, be installed, and join the domain; AI only needs a browser, a corporate card, or a click of a switch in software that has already been paid for.
2. The offboarding checklist is organized by “system,” while AI comes in by “identity”
Most companies' offboarding checklists look like this: email, ERP account, VPN, access card, WeCom, shared drive. This checklist is organized by "system"; its underlying assumption is that everything employees use is within the systems we manage.
AI breaks this assumption. It has four ways of coming in, and only the first is within IT's field of view.
| Entry method | Who approved it? | Will it stop when leaving the job | Where to stay |
|---|---|---|---|
| Enterprise seats centrally procured by IT | IT | Usually stops | IT Ledger |
| Employees register with personal accounts, then authorize reading company files | No one approved | Usually not | In the third-party account's “Authorized Apps” list |
| Personal subscriptions purchased with a company card for business use | Department Head | Finance has invoices, IT has no ledger | in the reimbursement form |
| The built-in AI switch in existing software | The administrator clicked once | Look at that software itself | Software backend toggle page |
There is also a fifth type, the easiest to forget: an agent or automation built by the company itself uses a dedicated service account. This account has no departure date, because it never had an onboarding date.
12
The number of AI tools a CIO thinks they have approved
152
In the same survey, the average number that large enterprises are running
The 60 days weren’t deliberately left by anyone; they emerged from several small things stacking up: third-party authorizations don’t expire just because you deactivated a company account; seats bought with a company card are paid annually, and unless finance actively cancels the subscription, they keep renewing; the vendor’s login state has a refresh mechanism—as long as that device is still on, the session can continue; and on top of that, HR, IT, and finance each manage one part, so no one has the full picture. Looked at individually, none of the four things is wrong; stacked together, they add up to 60 days.
A criterion for judgment: if something's permissions can only be disabled by “finding that person,” then it is not actually under control—it is just that nothing has gone wrong yet.
III. Deactivating an Account Does Not Equal Revoking Permissions
In systems like Odoo, "the person has left" is usually only half handled. Deactivating the user (archiving) is one step, but credentials are a separate matter. It is recommended to check the following points one by one; don't assume they will follow automatically.
| Points to verify | Why is it easy to miss | How to handle it |
|---|---|---|
| API keys and access tokens | It is a standalone credential table, tied to the username but not following the account status. | List by person and revoke item by item; after revocation, check the logs to see whether there are still calls. |
| Execution identity of scheduled tasks | A task runs under a certain user; even if that person leaves, the task keeps running. | Replace the execution identity with a dedicated account, and record it in the ledger |
| Email aliases and automatic forwarding | External emails continue to be delivered into the system, and the senders think it is still that person. | On the day of resignation, delete aliases, disable forwarding, and switch to a shared mailbox |
| Portal and External Accounts | Customer and supplier portal accounts are often treated as "customer data" rather than "employee permissions" | Clean up by contact, not just internal users |
| Shared account (info@, sales@) | Three people knew the password; two left and no one changed it. | Switch to personal accounts with aliases; for those that must be shared, change the password quarterly. |
| Dedicated account for the agent | For convenience, full permissions or superuser were granted. | Rebuild according to the principle of least privilege for tools; if the caller's permissions alone can be used, do not use a dedicated account. |
The last line is a new problem from the past two years. To meet deadlines, many projects create a "full-privilege service account" for agents or integrations, reasoning that this way there will be no errors caused by insufficient permissions. This account then becomes a door that never closes: the people have changed three times, and it is still open.
I've seen even cruder practices in projects—granting superuser permissions directly to automation scripts, on the grounds that 'it's read-only anyway.' The problem isn't what it does now, but what it can do later. Permissions don't become smaller just because the user follows the rules.
IV. Vendors Are Racing Ahead While Enterprises Are Catching Up on Old Debts
In the same week, there was another news item that formed a sharp contrast with this.
On November 18, Workday announced plans to acquire semantic layer company Honeydew. Its approach is to compile business definitions into governed queries—logic is moved out of prompts, and algorithms are set before requests reach the model. More critically: permissions follow the definitions, and invalid or unauthorized requests are blocked before query execution rather than flagged after the fact.
“Models should not improvise calculations behind salary figures.” — Gabe Monroy, Chief Technology Officer at Workday
On one side is "block it before the query is executed"; on the other is "still alive 60 days after departure." This gap is not a technical gap but a process gap: vendors treat permissions as an attribute of the system, while many enterprises treat permissions as a personnel procedure.
The characteristic of a procedure is that it happens at a certain point in time—once the signature is done, it ends. Permission, however, is not a one-time action; it is a status that must be confirmed once every day.
Separating responsibilities is also useful. HR manages the checklist—which items in the offboarding process require signatures; IT manages accounts—the user status in the system; security manages credentials—keys, tokens, service accounts; the business manages seats—whoever bought a subscription cancels it themselves. Dividing these four things among four people is more effective than “collective responsibility,” because each item can be checked individually.
Companies that do things fairly properly usually have a simple approach: a single table lists all non-human identities—service accounts, integration accounts, and agent accounts—each with an owner and an expiration date; resignation forms get an extra column, placed next to the access card; every quarter, ten people are sampled for a 'ghost permissions' audit. None of these three things requires buying a new system—only someone to take ownership.
V. Seven Things to Stop on the Day of Resignation
Turn all of the above into a checklist that can be posted on the wall, with HR and IT each taking half:
- Third-party authorizations in personal accounts. Before an employee leaves, have them open the "Authorized apps" section of their frequently used accounts and cancel, one by one, those related to the company. Only they themselves can do this step.
- AI seats issued by the company. The per-person billing sheet must reconcile with that month’s roster; any discrepancy is money paid for nothing.
- API keys and tokens. Revoke them one by one, not delete the user; check the call logs once within a week after revocation.
- The owner of the service account. Each service account must list two people: the current owner and the backup owner. Accounts without an owner are stopped immediately.
- Email aliases and automatic forwarding. Handle it the same day; don't wait for IT scheduling.
- Passwords for shared accounts. Change them whenever someone leaves; don't wait for the quarterly change.
- Credentials in agents and automation. Who created them, whose permissions they use, and whether they can be stopped with one click—these three questions need answers.
VI. Three Actions
First, conduct an AI asset inventory, and the scope should include "not purchased by IT." Three sources: IT procurement records, personal subscriptions in financial reimbursement forms, and the AI switches in all paid software. The count from the inventory will most likely make you rethink your own IT environment.
Second, add a field for “AI Authorization and Keys” to the offboarding form. Keep it together with the access card, and both parties in the handover must sign. If this field is not added, all the preceding governance actions lack a final closure.
Third, write an owner and an expiration date for each service account. When the expiration date arrives, automatically review it once: Is it still needed? Are the permissions still appropriate? This is more effective than an annual audit because it happens frequently enough.
People leave, but permissions do not leave on their own. This sounds like common sense, but the number 60 days shows that most enterprises still have not treated it as a rule that needs to be written into processes.
Sources: CloudEagle.ai, "State of AI Governance 2026" (2026-11-18, survey of 600+ CIOs/CISOs, vendor-commissioned), Workday press release (2026-11-18, proposed acquisition of Honeydew). The survey data reflect the commissioning party's framing and are not accompanied by an independent audit; Odoo-related verification items are based on general implementation experience, and actual behavior should be subject to testing with the version in use.
Zhongdi Information · Observations on Enterprise Management Digitalization · Data in this article are cited from public reports
