# Shadow AI at work: how do you bring it under control?

> Bringing shadow AI under control means writing a readable AI policy, providing an approved and properly hosted tool, and having a person review every high-stakes output. Train before you sanction.

## 01 The answer in plain terms

**Shadow AI** is the use, by employees, of artificial intelligence tools that the company has not chosen, approved or secured. An employee who pastes meeting minutes into a consumer assistant using a personal account is part of it. They do not mean any harm. They want to work faster.

Bringing it under control means setting rules of the game rather than banning it. Which tools are allowed, on which data, for which uses, and who reviews what. This guide is written for an executive committee that has to decide. It proposes a five-step method and rests on three non-negotiable measures: a written AI policy, an approved and properly hosted tool, and human validation of high-stakes outputs. One principle ties them together: train before you sanction.

The detailed survey data is in [the Shadow AI Barometer 2026](/en/glossary/shadow-ai-barometer-2026/). This page is the practical guide. It keeps only the figures that matter for the decision.

## 02 A practice that has become widespread

Generative AI has entered daily work, with or without the company's agreement. In the United States, 45.2% of working adults aged 18 to 64 use it for their work in Q2 2026, up from 37.4% in August 2025 (Federal Reserve Bank of St. Louis series, published 4 August 2026). In France, 73% of private-sector employees use AI at work, 10 points more than a year earlier. And 55% do so regularly or daily (Greenworking × Ipsos bva, 23 September 2026).

The same study measures the gap with official equipment. Only 37% of employees systematically use tools approved by their company. Usage is moving faster than tooling.

Part of this usage escapes the rules. The PagerDuty survey, run by Wakefield Research in April 2026 among 1,250 office workers outside IT (United States, United Kingdom, Australia, Japan, companies with $500M in revenue or more), gives a clear figure: 66% of office workers who use AI for their work have already done so with a tool they believed was not authorized (published 11 June 2026).

One signal points the right way. In the Netskope Threat Labs Report: Europe 2026 (25 May 2026), 43% of AI users go through personal accounts, down from 79% a year earlier. Accounts managed by the company rise from 28% to 72%. Moving to controlled accounts is therefore possible, and it is already under way.

WYP reading These figures come from different countries, panels and periods. Do not add them up. Together they say one thing: usage is established, and it comes before the official tool.

## 03 What really leaks

People readily imagine stolen trade secrets. The surveys describe the ordinary flow of work instead. In the PagerDuty study, 88% of users have already shared professional information with a public AI. According to the same survey, the content shared is:

- emails and correspondence, 43%
- meeting minutes, 40%
- customer data, 34%
- financial information, 31%
- confidential documents or strategies, 31%
- code or specifications, 23%
The Netskope report specifies the nature of the risk: 59% of data policy violations linked to AI and personal applications involve regulated data.

There is a cost as well. According to IBM (Cost of a Data Breach 2025, Ponemon Institute, 600 organizations, 30 July 2025), 1 in 5 organizations that suffered a data breach did so because of shadow AI. High shadow AI usage is associated with an average additional cost of $670,000.

There is a second, less visible risk. In the PagerDuty study, 38% of users handed in AI-assisted work without disclosing it. No one then knows which content a machine produced, or whether anyone reviewed it.

## 04 Why shadow AI appears

Four causes keep coming back. None of them is bad faith.

1
### The personal tool arrives before the company tool

In the PagerDuty study, 89% of users first adopted AI in their personal life. 79% now use it more at work than outside work. The habit is formed before the company has decided anything.

2
### The approved alternative is missing, or it disappoints

Only 37% of employees systematically use approved tools (Ipsos bva, 23 September 2026). When the official tool does not exist, arrives late or does less than the consumer tool, employees go elsewhere.

3
### The rule is absent or unreadable

63% of the organizations that suffered a breach have no AI governance policy or are in the process of writing one (IBM, 30 July 2025). Without a written rule, everyone sets their own.

4
### A pure ban moves the problem

Banning without offering an alternative does not remove the need. It only removes visibility. Leadership no longer sees what is happening, and teams stop talking about it. This point is an analysis, not a measured figure.

## 05 What it means legally

The question from an executive committee is direct: who answers when the AI gets it wrong? The Moffatt v. Air Canada ruling (2024 BCCRT 149, Civil Resolution Tribunal of British Columbia, 14 February 2024) gives an answer. A customer had been misinformed by the chatbot on the airline's website. The tribunal ordered Air Canada to compensate him, with CA$650.88 in damages, or CA$812.02 with interest and fees. Air Canada argued that its chatbot was a separate legal entity. The tribunal rejected that argument. A company answers for what the AI it makes available produces.

This ruling is not about shadow AI. It concerns a tool officially deployed by the company. The logic nonetheless extends to the subject of this guide. If a company answers for the AI it makes available, it will struggle to argue that it does not answer for the AI its employees use to reply to its customers. This reasoning is a reading, not an established rule. Have your legal counsel validate it.

Two other points come on top: the protection of personal data, as soon as customer or employee data is entered into an uncontrolled tool, and the duty of confidentiality toward your customers and partners. The European framework for AI (the AI Act) also expects the people who use AI to be trained. It is one more reason to train before sanctioning.

WYP reading The Air Canada ruling recalls a simple principle: the company signs what its AI produces. An unapproved tool, unreviewed results and no written rule leave it with little means to show that it took precautions.

## 06 The framing method in five steps

This method is steered at executive committee level, together with the IT department, legal, human resources and the business lines. It runs from measurement to training, in that order.

1
### Measure without sanctioning

Build a picture of actual usage. A short, anonymous survey of the teams, a cross-check with the IT department's technical data (applications visited, accounts used), a few interviews with managers. Say from the start that this is not a hunt for culprits. Without that promise, you will get polite answers and a false inventory.

2
### Classify data and uses

Three levels are enough. Data that can go into an approved AI tool, data that goes there only with precautions, data that never goes there. Tie each level to concrete examples from the business, otherwise the classification stays theoretical.

3
### Write the policy

A short document, readable in a few minutes. It states which tools are authorized, for which uses, with which data, who approves, and whom to ask in case of doubt. Date it, name an owner, have leadership stand behind it.

4
### Provide the approved tool

A policy without a tool is a ban. Choose a tool under a professional contract, with company-managed accounts, a known hosting location, a written commitment on how your data is used and controlled access rights. The detail is in the second non-negotiable measure.

5
### Train, then follow up

A short training per business line, using the company's real cases. A referent in each team to answer questions. A regular review of usage to adjust the policy. Sanctions come as a last resort, for a repeated breach after training. Training before sanctioning is not leniency. It is the condition for the rule to be understood, and therefore enforceable.

Five traps come back in framing efforts:

- banning across the board without offering a tool
- drafting a lawyer's policy that nobody reads
- choosing an approved tool that is worse than the consumer tool, then wondering why nobody uses it
- measuring once and never looking again
- leaving the subject to the IT department alone, when it also concerns legal, human resources and the business lines

## 07 The three non-negotiable measures

If your committee keeps only three decisions, these are the ones.

1
### A written AI policy

It sets the authorized tools, the data levels, the permitted and forbidden uses, the owner and the procedure in case of doubt. The IBM figure quoted above is a reminder: most of the organizations that suffered a breach had no AI governance policy, or were still writing one.

2
### An approved and properly hosted tool

Approved means signed off by leadership, the IT department and legal. Properly hosted means four checks: where the data is stored, whether it is used to train the model, who can access it, and what is logged. IBM finds that 13% of organizations suffered a breach of their AI models or applications, of which 97% lacked proper access controls. In the Netskope report, company-managed accounts rise from 28% to 72%. An approved tool only works if it is actually used.

3
### Human validation of high-stakes outputs

Define what a high-stakes output is: anything that goes to a customer, a regulator or a partner, anything that commits a price, a promise or a figure, any published text. An identified person reviews and signs. That is the Air Canada lesson. It is also the answer to the 38% of users who hand in AI-assisted work without disclosing it.

WYP reading The three measures hold together. The policy states the rule. The tool makes the rule easy to follow. Human review lets the company answer for what it publishes. Remove one, and the other two hold poorly.

## 08 What WYP does, and where to start

WYP.agency is a strategy consultancy for executives and executive committees. WYP does not sell AI tools, does not develop software and does not do marketing. On this subject, the work is to cross several AI models, cross-check their answers, identify blind spots and contradictions, and then a person reviews and signs. AI sharpens the decision, it does not make it.

[The AI Framing offer (Cadrage IA)](/en/cadrage-ia/) applies the method of this guide: audit of usage, written AI policy, approved tool, presentation to the executive committee. It is a fixed-scope package at €2,900 excl. VAT, delivered in 10 working days. WYP does not replace your legal counsel or your IT department. It frames the decision and provides the documents.

To start with no commitment, take [the AI maturity diagnostic](/en/ai-diagnostic/), free, in 90 seconds. For the terms used in this guide, the [AI glossary](/en/glossary/) defines them one by one. For the survey data, see [the Shadow AI Barometer 2026](/en/glossary/shadow-ai-barometer-2026/).

## FAQ

**What is shadow AI at work?**
It is the use of AI tools by employees without company approval, for example a consumer assistant used with a personal account. In the PagerDuty survey (11 June 2026), 66% of office workers who use AI for their work have already done so with a tool they believed was not authorized.

**Should companies ban consumer AI tools?**
Not without an alternative. A ban alone moves usage out of your sight. Provide an approved and properly hosted tool first, publish a written policy, then restrict what must be restricted. The Netskope report (25 May 2026) shows that moving to managed accounts is possible: they rise from 28% to 72%.

**Which data most often goes to public AI tools?**
According to the PagerDuty survey (11 June 2026), emails and correspondence (43%), meeting minutes (40%) and customer data (34%). Next come financial information (31%), confidential documents or strategies (31%) and code (23%).

**Is a company responsible for what an AI produces?**
In Moffatt v. Air Canada (2024 BCCRT 149, 14 February 2024), a Canadian tribunal rejected the argument that the chatbot was a separate legal entity and ordered the company to compensate a misinformed customer. A company answers for what the AI it makes available produces. The case concerns an official tool. Its reach for shadow AI is a reading to validate with your legal counsel.

**What should a company AI policy contain?**
The authorized tools, the data levels (what can go in, what goes in with precautions, what never goes in), the permitted and forbidden uses, the rule of human review for high-stakes outputs, the owner and the person to contact in case of doubt. A short, dated document backed by leadership.

**Should employees who use an unapproved AI be sanctioned?**
Not first. Only 37% of employees systematically use approved tools (Ipsos bva, 23 September 2026). A gap this widespread is not solved by sanctions. Provide the tool, publish the rule, train. Sanctions remain possible as a last resort, after training.

## See also

- [Shadow AI](https://wyp.agency/en/glossary/shadow-ai/)
- [Shadow AI Barometer 2026](https://wyp.agency/en/glossary/shadow-ai-barometer-2026/)
- [Guardrails](https://wyp.agency/en/glossary/guardrails/)
- [AI Act](https://wyp.agency/en/glossary/ai-act/)
- [Usage discipline](https://wyp.agency/en/glossary/usage-discipline/)
- [Auditing AI maturity](https://wyp.agency/en/glossary/ai-maturity-audit/)
- [Making generative AI legally safe](https://wyp.agency/en/glossary/secure-generative-ai/)

## Sources

- PagerDuty Shadow AI Survey, Wakefield Research, fieldwork 9 to 20 April 2026 (1,250 office workers outside IT, companies with $500M in revenue or more, United States, United Kingdom, Australia, Japan, margin of error ±2.8 points), published 11 June 2026 (https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/)
- Generative AI, Adoption Rate for Work: Employed Adults (series RPSGENAIUSAGESHAREWORK, Real-Time Population Survey, workers aged 18 to 64), FRED, Federal Reserve Bank of St. Louis, Q2 2026 value published 4 August 2026 (https://fred.stlouisfed.org/series/RPSGENAIUSAGESHAREWORK)
- Greenworking × Ipsos bva, “AI at work: behind mass adoption, ghost usage and managers on the front line”, 23 September 2026 (fieldwork 12 to 25 May 2026, 1,100 private-sector employees excluding blue-collar workers, including 474 managers) (https://www.ipsos.com/fr-fr/ia-au-travail-derriere-ladoption-massive-des-usages-fantomes-et-des-managers-en-premiere-ligne)
- Netskope Threat Labs Report: Europe 2026, 25 May 2026 (data from 1 March 2025 to 31 March 2026) (https://www.netskope.com/resources/threat-labs-reports/threat-labs-report-europe-2026)
- IBM, Cost of a Data Breach 2025 (Ponemon Institute, 600 organizations), 30 July 2025 (https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls)
- Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal (British Columbia), 14 February 2024 (https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/shadow-ai-framing-guide/ · Last updated : 2026-10-16