Notes on AI and leadership

Shadow AI: what it is and what leaders should do about it

What shadow AI is, why employees use AI tools nobody approved, what it costs and what a leadership team does in the next 30, 60 and 90 days.

Discuss a workshop

By Johannes Sundlo, AI and future of work advisor, formerly HR at Spotify. Updated 1 November 2026.

The short answer

Shadow AI is the use of AI tools at work that the organization has not approved, such as a personal ChatGPT, Claude or Gemini account or an AI browser extension. People rarely do it to break rules. The approved tool is missing or weak, the rules are unclear or nobody has shown them how. In my view, a ban mostly moves the use to personal phones. What works better is a 90-day plan. Find out what people use, approve a tool that does the job, write one page of rules, train people and measure.

See the 90-day plan
  • What AI tools used at work without the organization’s approval.
  • Why No good approved tool, unclear rules or no training.
  • What to do Map, approve, set rules, train, measure.

What is shadow AI?

Shadow AI is any AI tool or feature that employees use for work without the organization’s approval. The term comes from shadow IT, the older name for software that teams start using without asking IT.

What are examples of shadow AI?

Typical examples look like this. Someone pastes a customer email into a personal ChatGPT login to draft a reply. A recruiter uploads CVs to a free AI tool to get a shortlist. A manager installs an AI browser extension that reads everything on the screen. A team lets an AI note-taker nobody approved join customer calls. Someone runs a desktop AI agent with access to their work files.

What is the difference between shadow AI and shadow IT?

Shadow IT is any technology used without IT’s approval, like a team’s own file-sharing account. Shadow AI is the same for AI, with three differences. What people type in can end up with the vendor, and some consumer services train on it unless the user opts out. The answers feed reports and decisions, sometimes unchecked. And AI agents can act on their own with the employee’s access.

How common is shadow AI?

71%of 2,003 UK employees in a Microsoft study have used unapproved consumer AI tools at work
53%of HR professionals in a Swedish survey say people in HR use AI tools that are not formally approved
66%say so in the public sector, compared with 46% in the private sector
49%say it happens even where the organization has an AI policy or clear guidelines

In a Microsoft UK study from October 2025, 71 percent of 2,003 employees had used unapproved consumer AI tools at work (Microsoft UK). In Microsoft’s 2024 Work Trend Index, with 31,000 knowledge workers in 31 markets, 78 percent of people who use AI at work brought their own AI tools (Microsoft). That isn’t always against the rules, but it starts outside what the employer provides.

What does a survey of 329 HR professionals in Sweden show?

The AI & HR Barometern 2026, humAIn Academy’s survey of 329 HR professionals in Sweden, asked whether AI tools that are not formally approved are used in HR. 53 percent say yes, 32 percent say no and 16 percent don’t know. The share saying yes is 66 percent in the public sector, 46 percent in the private sector and 49 percent where there is an AI policy or clear guidelines.

Paying for tools doesn’t seem to change the share. It is 53 percent among the 289 respondents (88 percent) who name at least one AI tool their employer pays for, and the same among the 40 who name none or don’t know. It is lower where AI is part of the daily work. Among the 110 who describe AI in HR as recurring, integrated or strategically governed, 41 percent say AI tools that are not formally approved are used in HR, against 60 percent of the 200 still at individual tests or pilot projects. My reading is that a policy or a license doesn’t bring AI use into the open by itself. An approved tool built into the daily work has a better chance.

Source: AI & HR Barometern 2026, humAIn Academy, presented in State of AI in HR. 329 HR professionals in Sweden, 8 June to 24 September 2026. The question asks whether unapproved tools are used in HR, not whether the respondent uses them. Respondents opted in, subgroups are small and a correlation doesn’t show cause and effect.

Why do employees use AI tools that aren’t approved?

People use them to get the work done.

  • The approved tool is missing or weak. In Microsoft’s UK study, 28 percent said their company doesn’t provide a work-approved option.
  • The rules are unclear. If nobody has said what’s allowed, everyone makes their own call.
  • Nobody has shown them how. In the Work Trend Index, only 39 percent of AI users had received AI training from their company. Without it, people experiment in their own accounts.
  • It feels like cheating. In the same survey, 52 percent of AI users were reluctant to admit using it for their most important tasks. People don’t ask for approval for something they’d rather not talk about. Often it comes from a fear of looking replaceable. Here’s what leaders can do about AI anxiety.

All four are leadership problems. Blocking an app doesn’t give people a better tool, clear rules or training.

What are the risks of shadow AI?

  • Personal data. Pasting details about customers, candidates or colleagues into a personal AI account shares personal data with a vendor the organization has no agreement with. In the EU that is hard to square with the GDPR (EUR-Lex), and in the UK with the UK GDPR.
  • Confidential information. As of November 2026, OpenAI may use content from its services for individuals, such as ChatGPT, to train its models unless the user opts out. Business plans are excluded by default (OpenAI). So a contract or board memo pasted into a personal account can end up as training data.
  • Quality. AI writes convincingly even when it’s wrong, and without review the errors end up in reports and decisions.
  • The EU AI Act. Since an amendment entered into force on 27 July 2026, Article 4 requires providers and deployers of AI systems to take measures that support AI literacy among their staff and others who use AI on their behalf. It no longer demands a specific level, but the duty remains (EUR-Lex). From 2 December 2027, the high-risk rules in Annex III apply, including AI used in recruitment and decisions about workers. Neither is easy to meet for tools nobody knows about.
  • Cost. In IBM’s 2026 Cost of a Data Breach report, which studied 602 breached organizations, security incidents involving shadow AI more than doubled, to 43 percent from 20 percent the year before. Those breaches cost $5.39 million on average, against $4.99 million for breaches overall (IBM). And the licenses you pay for don’t deliver if people do their best work in personal tools.
  • AI agents. Desktop agents can read files, open email and act in other systems with the employee’s access. Claude Cowork is one of them, and my guide covers what happens to your data in Cowork, including on personal plans.

Does banning ChatGPT work?

Not on its own, in my view. A ban tends to move the use from the work laptop to the personal phone, and you lose what little visibility you had. A block can make sense for certain data, or briefly while you set up an approved alternative.

How should leaders manage shadow AI in the next 90 days?

The plan belongs to the leadership team. Bring IT and HR in, and give it one quarter.

  1. Days 1 to 30. Find out and approve. Ask teams which AI tools they use and for what, anonymously if needed, and make it clear nobody gets in trouble for answering. Ask IT what it can see. Then approve a tool. As of November 2026, Microsoft Copilot Chat is included with eligible Microsoft 365 licenses (Microsoft Learn). It comes with enterprise data protection, and prompts aren’t used to train the underlying models (Microsoft Learn). If it isn’t enough, compare the options in Copilot vs ChatGPT, or for HR tasks in Which AI tool is best for HR.
  2. Days 31 to 60. Set the rules and train. Write one page of rules. Train people on tasks from their own work, starting with the leadership team. The AI literacy playbook lays out the order.
  3. Days 61 to 90. Measure and adjust. Repeat the anonymous question and compare. Track use of the approved tool and the time it saves on a few common tasks. In the AI & HR Barometern 2026, 70 percent say they don’t measure the effect of AI in HR. Agree on a simple way to request a new tool, and review the approved list every quarter.

Talk to me about your planA few lines are enough to start.

What do you tell your team tomorrow?

A short message is enough to bring AI use into the open. Adapt this example to your situation.

Hi everyone,

Several of you already use AI at work, and that's good. Now let's do it in a way that protects us and our customers. From today, three rules apply.

1. Use [approved tool] for work tasks, signed in with your work account.
2. Never put personal data, customer data or confidential information into personal AI accounts.
3. Check what the AI wrote before you pass it on.

Using another tool that works well? Tell me, and we'll look at approving it. Nobody gets in trouble for having tried something.

[Name]

What should a shadow AI policy include?

Keep it to one page, so people read it.

  • Which AI tools are approved, and for what.
  • Which data never goes into AI, with examples such as CVs, salary data or customer contracts. My list of what to keep out of a prompt works as a starting point.
  • Who reviews AI output before it goes to a customer, a candidate or the board.
  • How to request a new tool, and how quickly you answer.
  • Who to ask when unsure.

For a broader governance checklist, see Ethical AI in HR. If you read Swedish, my AI policy page has a free template in Word. When one page is no longer enough, my AI governance framework adds an owner, decision rights, risk tiers and a committee charter.

How can IT detect shadow AI in Microsoft 365?

If you run Microsoft 365, these are the main tools for IT as of November 2026.

  • Defender for Cloud Apps. Its app catalog has a Generative AI category with more than a thousand AI apps. IT can see which ones are used, check the risk of each and sanction or block them (Microsoft Learn).
  • Microsoft Purview. Data Security Posture Management (DSPM), or the classic DSPM for AI, shows how AI is used and offers ready-made policies to protect sensitive data. On onboarded Windows computers, data loss prevention can warn or block people who paste sensitive information into AI sites (Microsoft Learn).
  • The Shadow AI page in the Microsoft 365 admin center. New, and in public preview through Microsoft’s Frontier program. It detects AI apps and agents on employees’ computers, such as ChatGPT Desktop, Claude Desktop, Ollama and OpenClaw. In the preview it can block only OpenClaw, and only on Windows devices managed in Intune. It requires opting in to Frontier, a Microsoft 365 E5 license, Defender for Endpoint and Intune (Microsoft Learn).

None of them reaches a personal phone or tells you why people chose the tool they did. For that, ask your team.

Frequently asked questions

What is shadow AI?

Shadow AI is the use of AI tools at work that the organization has not approved, such as personal ChatGPT, Claude or Gemini accounts or AI browser extensions. It is a form of shadow IT with bigger risks, because what people type in can leave the organization.

What is the difference between shadow AI and shadow IT?

Shadow IT is any technology used without IT’s approval. Shadow AI is the same for AI tools, but what people type in can end up with the vendor, the answers feed decisions without review and AI agents can act with the employee’s access.

Is shadow AI illegal?

Not in itself. But putting personal data into an AI service your employer has no agreement with can breach data protection law such as the GDPR, and it can break your employer’s rules, confidentiality duties or client contracts. Use the approved tool with your work account.

How common is shadow AI?

Very common. In a Microsoft UK study from October 2025, 71 percent of 2,003 employees had used unapproved consumer AI tools at work. In the AI & HR Barometern 2026, a survey of 329 HR professionals in Sweden, 53 percent say people in HR use AI tools that are not formally approved.

How do you stop shadow AI?

Not with a ban alone. Approve a tool that is good enough for the work, write one page of rules on which data never goes into AI, train people on real tasks and follow up every quarter. IT tools for discovery and blocking help, but they don’t reach personal phones.

Sources

Want to bring AI use out into the open?

In a half-day leadership workshop, you map where AI is already used, agree on the guardrails and draft a one-page policy. The leadership teams I’ve worked with give an average rating of 4.97 out of 5, measured across 13 groups.

Talk about your situation