58% of executives have had an AI incident. Nearly all are confident

Three numbers that cannot all be right

Okta, with Apprize360, surveyed nearly 300 technology executives and 500 knowledge workers, publishing on 28 May 2026. Take three of the findings together:

Okta shadow AI research, May 2026

Finding	Figure	Population
Using personal AI tools without approval	More than half	Employees
Using unsanctioned AI (US)	Two thirds, with nearly a quarter doing so regularly	Employees
Say organisational policy is unclear, hard to find, or non-existent	More than half	Employees
Had an AI-related security incident or close call in the last year	58%	Executives
Confident employees are using AI responsibly	Nearly all	Executives

The last two rows are the briefing. Fifty-eight per cent have already had an incident or a near miss, and nearly all of them are confident. Confidence at that level, in a population where a clear majority have had something go wrong, is not an assessment. It is the absence of one.

Why the policy-first reflex fails

The standard response to discovering shadow AI is to write an acceptable use policy. The employee data says why that does not work on its own: more than half already report that policy is unclear, hard to find, or absent. Publishing another document into that environment changes nothing measurable.

The access grants are the part that turns this from a compliance irritation into a security finding. This is not staff pasting a paragraph into a chatbot. It is staff connecting tools to internal messages, HR records and confidential documents, the same connector-authority problem as MCP tool access, arriving through a different door and with nobody in security aware it happened.

Record every system in one of three states

Inventory first, and the order matters

The sequence that actually produces compliance

1. SaaS and network telemetry, OAuth grantsMeasure what is in usethen
2. Client calls outrank public codeRank by data sensitivitythen
3. Give people a sanctioned routeLegitimise the safe majoritythen
4. Now it governs a mapped landscapeWrite the policy

The sequencing is not pedantry. A policy that bans the transcription tool a team already runs its client calls through produces one of two outcomes: the team stops disclosing, or the team ignores the policy. Both end with less visibility than before the policy existed.

Legitimising the safe majority is what makes the genuinely risky cases surface voluntarily, and voluntary disclosure is the only discovery mechanism that scales past what telemetry can see.

Where the systems actually are

The word shadow suggests concealment, and that framing sends people looking in the wrong places. Almost none of this is hidden. It is unrecorded, which is a different problem with a different solution.

Four places account for most of it, and each is discoverable without asking anybody a question.

Inside software you already pay for. A summarise button appeared in a release note. Nobody decided to adopt AI; a supplier shipped it. This is the largest category and the least visible, because the licence predates the feature.

On a departmental card. A subscription below the threshold that would trigger procurement review, expensed monthly, invisible to IT and perfectly visible to finance.

In a browser extension. Installed by one person, running against whatever page is open, including pages showing client data.

In a workflow somebody built. An automation calling a model API with a key created for a proof of concept that never ended.

The common thread is that every one of them leaves a record somewhere: a release note, a card statement, an extension inventory, an API key. The inventory is an exercise in reading systems you already have rather than in interviewing people, which is why it takes days rather than the quarter people budget for it.

The obligation nobody connects to this

The EU AI Act's AI literacy duty under Article 4 has applied since 2 February 2025 and was not touched by this year's deferral. It requires providers and deployers to ensure a sufficient level of AI literacy among staff dealing with AI systems.

An organisation where more than half of employees cannot find the AI policy is not in a strong position on that duty, and the duty has been live for eighteen months.

What to check this week

The AI policy generator produces the document once you have the inventory, the AI literacy tool covers the Article 4 duty, and the acceptable use policy tutorial walks the drafting.