This landed in the security press this week, and almost nobody is framing it the way it actually matters.
TrustedTech surveyed 2,000 workers in the UK and US on shadow AI use — the tools employees are running outside IT approval, outside corporate policy, outside any audit trail. The headline finding: 65% of decision-makers are using unapproved AI tools at work. That's more than double the rate of employees below the decision-maker level, at 31%.
The Register and Help Net Security both picked this up over the last three days. The number that stuck with us: 56% of those same decision-makers say they're concerned about their employees using shadow AI.
Let that sit for a second.
The governance paradox nobody says out loud
Here's the pattern playing out in organizations right now. A director or senior manager tells their team: "We need an AI policy. We need to know what tools you're using." Maybe they send a memo. Maybe they sit in on a vendor briefing about shadow AI risks.
Then they go back to their desk and ask ChatGPT to help draft the board presentation, summarize a sensitive meeting, run some analysis on quarterly financials, or pull competitive research on a partner. Through their personal account. No audit trail. No IT visibility.
Why? The TrustedTech data captures it directly: 69% of C-suite executives say speed trumps privacy or security. And 21% of shadow AI users are doing it specifically because they don't want their organization tracking their activity.
That last number is the one worth paying attention to. One in five shadow AI users — many of them senior — are using personal accounts as a deliberate move to stay off the institutional radar. Not because the approved tools are bad. Because they want to work outside the gaze.
Why small orgs get this worse, not better
In a 5,000-person organization, an executive who's shadow AI-ing their quarterly reports is one person inside a governance structure with layers. There's an IT department, a compliance function, maybe a dedicated AI committee. Bad behavior at the top can theoretically be caught and corrected.
In a 15-person nonprofit, the executive director running a personal ChatGPT account for donor strategy sessions IS the IT department, the compliance function, and the policy maker. There's no one above them to flag it. There's no internal escalation path.
And what leadership does, the team watches. If the ED is clearly using a personal AI tool — and people notice these things, fast — the implicit message is that the policy doesn't actually apply. The written rule becomes theater. Usage spreads, informally and invisibly, from the top down.
A second survey published this week from Okta corroborates the visibility gap: 96% of UK executives expressed confidence they know what AI tools their teams are using. Meanwhile, more than half of their workers are using unapproved tools. That gap doesn't exist because of reckless employees. It exists because no one actually checks, the tools are genuinely useful, and the person responsible for oversight is often doing the same thing.
What's getting processed with no controls
The category of work flowing through personal AI accounts at the leadership level is worth pausing on.
Board decks. Strategic plans. Budget scenarios. Funder relationship notes. Contract summaries. Personnel matters. In a nonprofit or public sector org, that likely includes confidential donor data, beneficiary information, and material covered by explicit grant confidentiality clauses.
This is not the same risk profile as an employee drafting a routine report with some light internal data. This is an organization's most sensitive strategic information going through a personal account, under consumer terms of service, with zero visibility into retention, training, or eventual exposure.
For a small business, that's competitive intelligence. For an NGO, that could be donor records and beneficiary data — material with explicit legal protections under PIPEDA, GDPR if you have any EU connection, and in many cases direct confidentiality clauses in grant agreements. For a public sector organization, some of that material has statutory protections.
"I was just trying to move faster" is not a defense in an audit or a breach notification.
Why the behavior keeps accelerating
The tools keep getting better and more compelling. A senior manager who tried AI tools two years ago and found them useful now has access to systems that can replace hours of skilled analytical work in minutes. The productivity argument gets stronger every quarter.
Meanwhile, corporate IT procurement moves slowly. Approved AI options consistently lag behind what's available on a personal account. The gap between what someone can do through their own ChatGPT or Gemini subscription versus what they're permitted to do through official channels has never been wider.
So shadow AI use makes rational sense at the individual level — fast, useful, frictionless — even as it creates real systemic risk at the organizational level.
The standard response — a policy memo, a training module, a compliance briefing — doesn't work. 77% of shadow AI users already acknowledge they're taking a security or privacy risk. They know. They continue anyway. Because knowing isn't the barrier; convenience and speed are.
What actually changes behavior
Organizations that are actually closing this gap at the leadership level are doing three things differently.
Start at the top, visibly. In a small org, if the founder or ED switches to approved tools and can articulate why — "I stopped using my personal account for work, here's what I use now" — it moves the culture faster than any written policy. People follow behavior, not documents. If leadership is still on personal accounts, the policy is aspirational at best.
Eliminate the "I don't want to be tracked" motivation. Twenty-one percent of shadow AI use is a deliberate self-protection move. Self-hosted tools remove that dynamic entirely. Local models running on infrastructure you control — Ollama with Open WebUI on a dedicated machine, or a private cloud deployment with tight access controls — mean data doesn't leave your environment. No third-party terms. No external retention. No concern about organizational visibility, because you control the environment and can set visibility rules that make sense. This isn't a month-long infrastructure project anymore. It's a sprint.
Make approved tools faster than the workaround. The reason people use personal accounts is zero friction — they're already logged in, they work, they're familiar. An approved shared workspace that's genuinely integrated into existing workflows, doesn't require a separate login, and covers the actual tasks people do will displace the personal accounts not because of policy enforcement but because it's the easier path. Friction reduction beats compliance pressure every time.
The bottom line
The TrustedTech data isn't a story about irresponsible employees. It's a story about an accountability gap that runs from the top of every organization — and is particularly acute in small orgs where the person who should be modeling the behavior and the person most likely to be breaking the policy are the same individual.
If you're about to brief your team on your AI governance policy, ask yourself first: when did you last use a personal AI account for something work-related? What went through it? If that question lands uncomfortably, your policy has a credibility problem before it's published.
The fix isn't more governance documents. It's the right tools at the right level, starting at the top, built on infrastructure people will actually use instead of work around.
We help small org leadership teams set up approved AI infrastructure — self-hosted when data sensitivity requires it, integrated into real workflows so people actually use it. Reach out if you want to talk through what that looks like for your org.