What it is
Find out which AI tools your staff already use for work, before you write any rule about them. Ask them directly, and promise that no one will be blamed for what they say. Then pick one approved tool for the tasks most people use AI for, and set up a quick way to ask for another one.
Shadow AI is the name for AI tools staff use for work without the company’s approval: a free chatbot on a personal account, a browser extension, a meeting notetaker someone signed up for. The problem is not that staff use AI. The problem is that nobody knows what goes into those tools, or who checks what comes out.
What the evidence shows
Three documented cases and two studies bear on this. No study has tested whether an amnesty or an approved tool reduces the use of unapproved ones.
- Staff use AI before the company decides anything. In a survey that Microsoft, which sells AI tools for work, ran with LinkedIn in early 2024, 78% of desk workers who used AI said they used tools their employer did not provide. The figure is checked here. It counts tools the employer did not provide, not tools it banned. A widely repeated MIT figure, that 90% of employees use personal AI tools, is a share of companies in a small sample, not of employees. That figure is checked here.
- Personal use runs ahead of company use. A survey of 179 leaders in Poland found that 83% used AI every day, while about 70% of their organizations had not moved beyond pilots.
- The company may not know until something goes wrong. At Victoria’s child protection service in Australia, about 900 staff, almost 13% of the department, had opened ChatGPT before any rule existed. The department found out after a worker typed a family’s details into it. The privacy regulator said the department had “almost no visibility” of how staff used the tools.
- One upload can expose thousands of people. At a state agency in Australia, a temporary worker uploaded a spreadsheet to ChatGPT, which the agency had not approved. It held personal and health details of 2,031 people. An often repeated IBM figure says shadow AI adds $670,000 to the cost of a data breach. It compares breached organizations with high and low levels of shadow AI, without allowing for other differences. The same report puts the extra cost of breaches that involved shadow AI at about $200,000. Both figures are checked here.
- A rule that requires approval is ignored when asking is slower than not asking. At the US law firm Butler Snow, the policy required written approval before using ChatGPT for legal research. A partner used it without approval and filed five made-up citations.
- People have a reason to use AI quietly. In four experiments with 4,439 people, workers expected colleagues to judge them as less competent if they used AI, and evaluators did judge them that way. Someone who fears that judgement will not tell a manager which tool they used.
How to do it
- Ask, with a promise of no blame. Send a short anonymous survey: which AI tools do you use for work, for which tasks, on a personal or a company account, and what do you put into them. Say in writing that no one will be disciplined for past use.
- Check what you can see. Ask your IT supplier for the AI sites and apps reached from company devices over the last three months, and the AI features switched on in software you already pay for. Logs show which sites were opened, not what was typed in, so they tell you where to ask.
- List what you find. One line per tool: who uses it, for what, and what data goes in. This list is the start of your rule on what data may go into which tools.
- Approve one tool for the common tasks. Take the tasks most people named, such as drafting, summarising and translating, and pay for one tool that covers them on a business plan. Check what the supplier does with what staff type in.
- Make asking quick. Name one person who answers requests for a new tool within a week. A request that waits a month is a request people stop making.
- Block only where a mistake would be serious. For teams that handle health, children’s, financial or legal files, block public AI tools and uploads to them. Elsewhere, a block without an approved tool moves the use to personal phones, where nobody sees it.
- Ask again in six months. Run the same survey. Put the result and the tool list in your one-page AI-use policy.
What to measure
- The share of staff who use AI for work only through approved tools, from the survey.
- The number of AI tools in use that are not on the list, from the survey and the IT check.
- How long a request for a new tool waits for an answer.
Where it fails
- The survey is not trusted. If people think the answers will be used against them, they will not answer honestly. Keep it anonymous and keep the no-blame promise.
- The approved tool is worse than the free one. Staff will go back to the tool that does the job. Ask the people who use AI most which tool to approve.
- The check covers company devices only. Use on a personal phone does not show in company logs. Only the survey reaches it.
- AI arrives inside software you already pay for. A meeting notetaker or an email assistant switched on by default is unapproved use that nobody chose. Check those settings, not only the websites.
- Temporary staff and contractors are left out. The upload at the Australian agency was made by a temporary worker. Include them in the survey and in the approved tool.
By company size
- Micro and small: ask everyone in one meeting which tools they use. Buy a business plan for the tool most people already use, and agree one rule: no client or personal details in any free tool.
- Medium: the anonymous survey, one approved tool for everyone, and one named person who answers requests for others.
- Mid-market and enterprise: the survey and the IT check together, a list of tools kept by the person who owns AI, and blocks for the teams with the most sensitive data.
Read next
- Butler Snow: a firm with an AI policy, and a partner who did not follow itThe firm’s policy required written approval before using ChatGPT for legal research. A partner used it without approval and filed five citations it had made up. The court removed three lawyers from the case. The firm paid for a check of 2,400 citations.
- “MIT Sloan shows that the barriers to digital transformation are mainly cultural”The page cited is a collection of short articles with no single author or date. It says cultural, organizational and technical shifts are all needed. It does not say that culture is the main barrier.
- Decide what data may go into which AI toolsSort your data into two or three groups and say which AI tools each group may go into. Client secrets and personal data stay out of public tools. Where a mistake would be serious, block the upload instead of relying on a rule.
Company cases
What happened without it
- Butler Snow The policy required written approval before using ChatGPT for legal research. The partner used it without asking for approval.
- Department of Families, Fairness and Housing (Victoria) About 900 staff, almost 13% of the department, had opened ChatGPT before any rule existed. The regulator found the department had “almost no visibility” of how staff used such tools.
- NSW Reconstruction Authority A temporary staff member uploaded the spreadsheet to ChatGPT, a tool the agency had not approved.
Sources
- Wojewodzic (2026). AI Transformation in Polish Organisations: The Gap Between Individual Adoption and Organisational Maturity. e-mentor, 2(114), 75–83. What this study did and found
- Reif, Larrick and Soll (2025). Evidence of a social evaluation penalty for using AI. Proceedings of the National Academy of Sciences, 122(19), e2426766122. What this study did and found
- Office of the Victorian Information Commissioner (September 2024). Investigation into the use of ChatGPT by a Child Protection worker. ovic.vic.gov.au
- Microsoft and LinkedIn (8 May 2024). 2024 Work Trend Index Annual Report: AI at Work Is Here. Now Comes the Hard Part. microsoft.com
Claims on this topic, checked against these studies:
Related practices
- Write a one-page AI-use policy staff will actually read Put the rules for AI on one page: which tools are allowed, what must never be typed in, who checks the output, and who to ask. Then build the most important rule into the work itself, because a page alone gets ignored. Do first · Effort: low (days, one person)
- Start with ready-made AI in software you already use, before building your own Your office, accounting, support and sales software probably has AI features now. Try those on one task, with a few licences and a measured result, before you pay anyone to build something custom. Do first · Effort: low (days, one person)
- Protect psychological safety during an AI rollout AI projects can make staff afraid to speak up and unwilling to share what they know. Say early what AI means for jobs, and consult the people affected. Do first · Effort: medium (weeks, a few people)