Shadow AI Agents Aren’t a Junior-Employee Problem. Your C-Suite Is Fine With Them.

Two-thirds of office professionals — 66% — have used an AI tool their employer never approved, according to a PagerDuty survey of 1,250 professionals at companies with over $500M in revenue, published June 2026. That number alone isn’t new; “shadow IT” has existed for twenty years. What’s new, and what should worry a CISO more than the headline stat, is who’s comfortable with it: a BlackFog survey of 2,000 UK/US employees (Sapio Research, published Jan-2026) found 69% of presidents and C-suite members, and 66% of directors and senior VPs, say they’re comfortable with unsanctioned AI use — prioritising speed over governance.

This isn’t rogue analysts hiding a ChatGPT tab from IT. It’s leadership deciding, explicitly or by not asking, that the risk is worth it.

Myth: “Shadow AI” means someone pasted a document into ChatGPT

That was the 2023 version. The 2026 version has hands. DNSFilter’s “Visibility Deficit” report (CensusWide, 400 IT/security professionals, US/Canada, June 2026) found 46% of organisations have had an employee connect an unauthorized AI tool directly to corporate systems — not copy-paste, a live connection with API or OAuth access. 43% of organisations let employees grant that access themselves, with IT finding out after the fact; at 6.5% of organisations, IT never finds out at all.

Once the tool has API access and can take actions on its own, it isn’t a chatbot anymore — it’s an unmanaged agent with a login. The same DNSFilter report ties this directly to outcomes: 20.1% of respondents said an unauthorized AI tool connection had already caused a security incident at their organisation.

What’s actually being shared

PagerDuty’s survey found 88% of respondents had shared work-related information with public AI tools: 34% shared customer data, 31% shared financial information or confidential strategy documents, 43% shared emails or correspondence. BlackFog’s numbers point the same direction from a different angle: 33% of employees have shared research or data sets, 27% have shared employee data (names, payroll, performance), 23% have shared financial statements or sales data — and only 53% say they understand how the data they enter gets stored or reused.

Neither survey is describing malice. BlackFog found 60% of respondents think the risk is worth taking if it helps them hit a deadline, and PagerDuty found 77% feel official AI restrictions are holding back their career development. People aren’t hiding this because they think it’s wrong. They’re doing it because the sanctioned alternative is slower, and leadership isn’t stopping them.

Why “just ban it” doesn’t work

Banning unsanctioned AI tools is the instinctive first move for a lot of security teams, and it’s the same mistake shadow IT made with personal Dropbox and Slack a decade ago: a ban doesn’t remove the demand, it removes your visibility into it. If 69% of your own C-suite is comfortable with unsanctioned use, a policy with no enforcement behind it just teaches everyone else to stop mentioning it. The 6.5% of organisations where IT never finds out about an unauthorized connection are not the ones with the strictest policy on paper — they’re the ones where nobody is actually watching for it.

Five signs you already have shadow agents, not just shadow chat

• A team ships faster than its approved toolset should allow. If a 3-person team is producing output that would normally need twice the headcount, ask what’s doing the extra work.
• An MCP server or API integration appears in your logs that nobody filed a change ticket for. See our MCP gateway governance guide for what a reviewed, sanctioned MCP setup looks like versus one nobody owns.
• A vendor tool your team adopted “to try it out” six months ago is still running and nobody remembers approving it permanently. Trials rarely get decommissioned; they get load-bearing.
• Your help desk has fielded a password-reset or access request for a tool that isn’t on the approved list. That’s an agent with a login IT didn’t issue.
• Nobody can tell you, today, exactly which systems your AI tools can reach. If the honest answer is “we’d have to check,” you have the same visibility gap DNSFilter measured at 46% of organisations.

What actually reduces the risk

Not a ban — a fast, sanctioned path that’s faster than going around IT, plus real visibility into what’s already connected. That’s the same logic as least-privilege AI agent permissions and agent observability we’ve written about for agents IT deployed on purpose — it applies with more urgency to the ones nobody deployed on purpose. And it’s the same containment principle Microsoft’s CEO argued for at treating every AI agent as an insider risk: the agent you don’t know about is the one you can’t contain. A fast audit of what’s actually connected — not what’s on the approved-vendor list — is the starting point, not the policy memo.

FAQ

What’s the difference between shadow IT and shadow AI?

Shadow IT is any unapproved software connected to corporate systems — personal cloud storage, unsanctioned SaaS. Shadow AI is the same problem with tools that can read, write and act on data autonomously once connected, which raises the stakes from “unapproved file sync” to “unmonitored agent with API access.”

Is shadow AI mostly a junior-employee problem?

No. BlackFog’s survey found C-suite and senior leadership are more comfortable with unsanctioned AI use than middle management, not less — 69% of presidents/C-suite and 66% of directors/senior VPs said so, versus wider hesitation further down.

How common is it for an unauthorized AI tool to actually cause an incident?

DNSFilter’s June 2026 survey of 400 IT/security professionals found 20.1% had already experienced a security incident tied to an unauthorized AI tool connection.

Does banning AI tools solve the problem?

It reduces visibility more than it reduces use. The fix that works is a sanctioned path fast enough that people don’t go around it, plus monitoring for connections nobody approved — the same logic used against other forms of shadow IT for two decades.

What should the first step be?

An honest inventory: what AI tools and agents are already connected to your systems, who approved them, and what they can reach. Most organisations can’t currently answer that — which is the risk itself, not a detail about it.

That inventory question gets sharper in regulated markets. Gulf regulators aren’t waiting for an AI law to act on it: in June 2026 the UAE fined a bank’s own compliance officer personally over a sanctions-screening failure, not the vendor behind the tool — see why the Gulf fines AI-era compliance failures without an AI law.

Sources

• Morningstar/Business Wire: PagerDuty Report Finds Two-Thirds (66%) of Office Professionals Have Used Unauthorized AI Tools at Work
• BlackFog Research: Shadow AI Threat Grows Inside Enterprises
• DNSFilter: Shadow AI Is Already Causing Security Incidents
• CIO.com: Roughly half of employees are using unsanctioned AI tools, and enterprise leaders are major culprits

Find out what’s actually connected to your systems — free audit.