# Shadow AI: What Should Leaders Do When Staff Use Unapproved Tools?

Canonical URL: https://blog.theaieducator.io/posts/what-to-do-about-shadow-ai
Publication: Dan Fitzpatrick Insights
Author: Dan Fitzpatrick
Topic: AI Adoption and Organisational Change
Published: 2026-09-12T07:16:00.000Z
Modified: 2026-09-12T07:16:02.622Z

Two thirds of office professionals have used AI at work believing they were not permitted to. Blocking it first takes away your visibility without changing the behavior. A way to diagnose what you find, and the three decisions that follow.

## In brief

Shadow AI, the use of AI tools an organization has not approved, is now the normal way many staff use AI at work. Leaders should not block it first. Diagnose what job the person was trying to get done and what the approved route took from them, then adopt the tool, route the work to an approved alternative, or stop it and say why. Restriction without a faster approved replacement moves the work onto personal accounts, where it cannot be seen, audited or improved.

## Key takeaways

- Shadow AI is any use of an AI tool at work the organization has not approved, and on current evidence it is the normal way many staff use AI rather than a fringe behavior.
- PagerDuty's June 2026 survey of 1,250 office professionals found two thirds had used AI at work while believing they were not permitted to, and 88 percent had put work information into a public AI platform.
- UpGuard found executives had the highest rate of regular unapproved use, so shadow AI is not confined to staff furthest from the policy.
- Understanding the policy did not reduce unapproved use in the UpGuard data and was associated with more of it, which means training alone will not close the gap.
- Blocking without a replacement moves the work onto personal accounts rather than stopping it: UpGuard found 45 percent of workers finding workarounds to reach blocked applications.
- Dan Fitzpatrick's Detour Question asks what job the person was trying to do, what the approved route took from them, and whether the leader would be happy had they complied, turning a list of violations into a list of jobs.
- Every piece of shadow AI resolves into one of three decisions: adopt the tool, route the work to a faster approved alternative, or stop it and say what to do instead.

Somewhere in your organization this week, someone did good work with a tool nobody approved. They did not tell anyone. They are not going to.

The instinct is to treat that as a breach. I think it is better read as evidence, and it is the most honest adoption data you will get all year.

So, the short answer. Do not block it first. Find out what job the person was trying to get done, decide whether that job is legitimate, then adopt the tool, route the work to an approved alternative, or stop it and say why. A block placed before that diagnosis takes away your visibility and leaves the behavior exactly where it was.

## What is shadow AI, and how common is it?

Shadow AI is any use of an AI tool at work that the organization has not approved: a teacher pasting a draft report into a chatbot on a personal account, a finance officer summarizing a supplier contract in a free web tool, a deputy head writing a parent letter on her phone on a Sunday evening. On the evidence available, shadow AI is not fringe behavior. It is the normal kind.

[PagerDuty's survey of 1,250 office professionals](https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/), run by Wakefield Research across the United States, the United Kingdom, Australia and Japan and published in June 2026, found that two thirds had used AI tools at work while believing they were not permitted to. Eighty-eight percent had put work information into a public AI platform. A third had put customer data in.

[UpGuard's State of Shadow AI](https://www.upguard.com/resources/the-state-of-shadow-ai), released in November 2025 and drawing on 2024 surveys of 1,500 security leaders and employees across seven countries, found more than 80 percent using unapproved tools. The detail that should stop a leadership team is who. As [Cybersecurity Dive reported](https://www.cybersecuritydive.com/news/shadow-ai-employee-trust-upguard/805280/), mid-level managers and frontline staff showed the highest overall use, and executives showed the highest rate of regular use. The people most likely to be reaching for an unapproved tool week in and week out are the ones sitting in the room where the policy was agreed.

## Why blocking it first makes the problem invisible rather than smaller

Blocking changes where the behavior happens, not whether it happens.

UpGuard found 45 percent of workers finding workarounds to reach blocked applications. In the PagerDuty survey, 77 percent said their organization's AI restrictions held back their professional growth, which tells you how a block feels from the far side of it. Neither number makes restriction wrong. Both say that restriction without a replacement pushes the work off your network and onto a personal account, where you can no longer see it, audit it or improve it.

The most uncomfortable finding in the UpGuard data is about awareness. Fewer than half of workers said they understood their organization's AI policy, and among those who said they did understand its security requirements, regular use of unapproved tools was more common, not less. I read that as a warning about a theory of change many leadership teams are still running on: that people break the rule because they do not know the rule. Some do. But a sizeable group knows the rule, has weighed it against the work in front of them, and has decided the work matters more. No amount of training reaches that group. Only a better approved route does. This is the point at which [AI adoption stops being a training problem and becomes a leadership problem](https://blog.theaieducator.io/posts/why-ai-adoption-is-really-a-leadership-problem).

## What the guidance gap has to do with it

Most people using AI without approval were never given an approved way to do the job.

A Gallup and Walton Family Foundation survey of 2,069 teachers, fielded in February and March 2026 and [reported by Education Week](https://www.edweek.org/technology/teachers-say-lack-of-ai-guidance-is-a-major-problem/2026/05), found that 82 percent had received no formal guidance on applying AI tools to their work, and that 34 percent had received no guidance of any kind. That is not a workforce ignoring instructions. It is a workforce that was never given any.

The same vacuum shows at the policy level. When the Commons Education Select Committee took evidence in September 2026, [Tes reported](https://www.tes.com/magazine/news/general/schools-desperate-clearer-ai-guidance) NASUWT's Darren Northcott describing existing Department for Education guidance as "quite laissez-faire" and asking for "much clearer rules of the road," while Pip Sanderson of the National Institute of Teaching summed up what schools say back: "This is all great and really interesting, but what do I do?"

Hold those two findings together. Most staff have no approved route. Most leaders have no clear rule to build one from. Shadow AI is what grows in that gap, and it grows fastest where the work is heaviest.

## The Detour Question: how to read an unapproved tool

A detour tells you two things: where people were trying to get to, and what was wrong with the road you built.

The Detour Question is what I ask about any unapproved use of AI before deciding what to do about it: What job was this person trying to get done? What did our approved way of doing it take from them, in time, in quality or in dignity? And would I be happy if they had done it the way we told them to? A leadership team that can answer all three has found a gap in its own provision. One that cannot has found someone to discipline.

The third question is the one that changes meetings. Take a hypothetical that will not feel hypothetical to anyone who has run a school: a head of department pastes a set of pupil comments into a free chatbot at eleven at night to turn them into report card sentences. The approved way was to write ninety of them by hand. Answer the third question honestly and the conversation stops being about the tool.

This is my suggested way of reading the behavior, not a tested method. But it does something a security audit cannot: it converts a list of violations into a list of jobs, and jobs are the thing a leadership team can actually act on.

## What I tell leadership teams

When I speak to leadership audiences internationally, the shadow AI question almost never arrives as a question about security. It arrives in the corridor afterward, quietly, as a confession. Someone has worked out what their team is already doing and does not know whether to say anything.

What I tell them is this. You have one asset here and it has a short shelf life: people are still willing to tell you. Every week spent deciding on a sanction drains some of that willingness away, and once it has gone you are not running an organization with a policy. You are running one with a rumor mill.

So the first move is not a rule. It is a window. Name a period, four weeks is enough, in which anyone can say what they are using and what they use it for, with nothing attached. Say plainly that nothing declared in that window will be held against anyone. Then keep that promise in the one case where keeping it is hard, because that is the case everyone will be watching.

Across the leadership teams I work with, what comes back is rarely a long list of rogue tools. It is a short list of the same three or four jobs, done slightly differently by people who each believed they were the only one doing it. That list is a gift. It is an adoption roadmap written by the people who will have to live with it. The staff running ahead of you need a different response from [How to Bring the AI Sceptics With You](https://blog.theaieducator.io/posts/how-to-bring-the-ai-sceptics-with-you), and a leadership team usually has both groups in the same building.

## Adopt, route or stop: deciding what to do with each one

Once you know the job, every piece of shadow AI resolves into one of three decisions.

| Decision | When it applies | What the leadership team owes people |
|---|---|---|
| **Adopt** | The job is legitimate and the tool is acceptable | Put it on the approved list in writing, name who owns it, and say who asked for it |
| **Route** | The job is legitimate, the tool is not | Provide an approved way to do that same job, and make it the faster one |
| **Stop** | The job should not be done this way by anyone | Say so plainly, say what to do instead, and say what happens if it continues |

Route is the one most organizations skip, and skipping it is why bans decay. If you take away the eleven o'clock chatbot and hand back the ninety handwritten comments, you have not made a decision. You have restored a problem. The approved alternative has to be faster than the detour, or the detour comes back within the term. Making that alternative stick is the same work as [moving from AI experimentation to organization-wide adoption](https://blog.theaieducator.io/posts/how-to-scale-ai-adoption-across-an-organization): it has to transfer from the person to the process.

Adopt has a discipline of its own. A tool that goes on the list needs the standard set out in [what schools should require before approving an AI tool](https://blog.theaieducator.io/posts/approving-ai-tools-for-schools): a written promise from the supplier, and a written statement from you about what it is for and what would take it off the list. Amnesty is not the same as approval.

## When a hard block is the right call

Some uses should be blocked immediately, without a diagnosis and without a window.

Anything that puts identifiable information about children, patients, staff records or safeguarding material into a consumer tool goes straight to stop. So does anything where a machine is making a decision about a person rather than drafting something for one. The Detour Question is for the large middle ground of ordinary work, not for the small set of uses where the answer is already settled. Say which category is which on the same day you open the window, or the window itself will be read as permission.

There is also an honest limit to the evidence here. The PagerDuty and UpGuard samples are corporate, not educational, and UpGuard's underlying surveys were run in 2024. The Gallup teacher data is US. I have not seen a study that measures unapproved AI use inside schools with the same rigor, and that absence is itself worth naming when someone quotes a percentage at you in a governors' meeting.

## Where this leaves your next meeting

Shadow AI is not a measure of how badly your people behave. It is a measure of how far your approved route has fallen behind the work they are actually being asked to do. Treated as misconduct, it teaches people to hide. Treated as a detour, it tells you exactly where to lay the next stretch of road.

If your leadership team is working through what staff are already doing with AI and what to approve next, this is the kind of work I support through [adoption and implementation programs](https://www.theaieducator.io/project-momentum) with schools and organizations.

## Sources and further reading

- PagerDuty, "Two-Thirds of Office Professionals Have Used Unauthorized AI Tools at Work," survey of 1,250 office professionals conducted by Wakefield Research, June 11, 2026. https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/
- UpGuard, *The State of Shadow AI*, released November 2025, drawing on 2024 surveys of 1,500 security leaders and employees. https://www.upguard.com/resources/the-state-of-shadow-ai
- Cybersecurity Dive, "Shadow AI is widespread and executives use it the most," November 12, 2025. https://www.cybersecuritydive.com/news/shadow-ai-employee-trust-upguard/805280/
- Education Week, "Teachers Say Lack of AI Guidance Is a Major Problem," on a Gallup and Walton Family Foundation survey of 2,069 teachers fielded February and March 2026, published May 28, 2026. https://www.edweek.org/technology/teachers-say-lack-of-ai-guidance-is-a-major-problem/2026/05
- Tes, "Schools 'desperate' for clearer AI guidance," reporting the Commons Education Select Committee, September 9, 2026. https://www.tes.com/magazine/news/general/schools-desperate-clearer-ai-guidance

*Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and an international keynote speaker who works with schools and organizations on leading through AI. [More about Dan](https://www.theaieducator.io/about).*


## Frequently asked questions

### What is shadow AI?

Shadow AI is any use of an AI tool at work that the organization has not approved, such as a teacher pasting a draft report into a chatbot on a personal account. It is distinct from approved use because nobody has checked what happens to the data or the work.

### How common is shadow AI in the workplace?

Very common. PagerDuty's June 2026 survey of 1,250 office professionals found two thirds had used AI at work while believing they were not permitted to. UpGuard's report, released in November 2025, found more than 80 percent of workers using unapproved tools across seven countries.

### Should we just block unapproved AI tools?

Blocking without an approved replacement changes where the behavior happens, not whether it happens. UpGuard found 45 percent of workers finding workarounds to reach blocked applications. Block immediately where children's or staff data is involved, and elsewhere provide a faster approved route first.

### Why do staff use AI tools the school has not approved?

Usually because no approved way to do the job exists. A Gallup and Walton Family Foundation survey of 2,069 teachers, fielded in early 2026, found 82 percent had received no formal guidance on applying AI to their work, and 34 percent had received none at all.

### How do we find out what staff are actually using?

Open a declaration window rather than an investigation. Name a period, four weeks is enough, in which anyone can say what they use and what they use it for, with nothing attached, and keep that promise in the hardest case, because everyone will be watching it.

### Is shadow AI a security problem or a leadership problem?

Both, but treating it only as a security problem misreads it. The pattern of use shows where an organization's approved route has fallen behind the work people are asked to do, which makes it adoption evidence that a leadership team can act on.

---
Source: [Shadow AI: What Should Leaders Do When Staff Use Unapproved Tools?](https://blog.theaieducator.io/posts/what-to-do-about-shadow-ai)
Publisher: [The AI Educator](https://theaieducator.io)