A global study of more than 48,000 people found that employees whose organization had banned generative AI reported uploading sensitive company information into public AI tools at 67%. Where no policy existed at all, the rate was 33%. The ban is associated with roughly twice the exposure of saying nothing. If you are trying to stop shadow AI, the prohibition is not a neutral starting position you can refine later. It is the thing to replace.
That finding comes from the KPMG and University of Melbourne study Trust, attitudes and use of AI, fielded across 47 countries. Its authors state the conclusion plainly: outright bans may be ineffective, and having a policy does not guarantee compliance. What closes the gap is guidance and education on responsible use.
This article is for the person at a US company of 50 to 500 people who has an AI prohibition on the books, suspects it is ignored, and has no security function to hand it to.
The ban tracks the behavior it forbids
Start with how people reach these tools. In the same study, 70% of employees who use AI at work said they do it through free, publicly available tools. Only 42% said they use tools their employer provides, and 18% pay out of their own pocket for a public tool. The default access path is a consumer login, and the employer-provided path reaches fewer than half the people already doing the work.
Then the policy layer, as employees report it: 41% say their organization has a policy or guidance on generative AI use, 6% say it has a policy banning generative AI, 34% say there is nothing, and 19% do not know. Prohibition is a minority strategy, and it performs worse than the majority one.
Across the whole sample, 48% to 49% of employees said they had uploaded sensitive company information, including financial, sales or customer data, into public AI tools. Fifty-six percent had used AI tools at work without knowing whether it was allowed.
Split that upload behavior by the policy the employee works under and the ordering is uncomfortable:
| Employee reports their organization has | Uploaded sensitive company information |
|---|---|
| Banned generative AI | 67% |
| A policy guiding generative AI use | 56% |
| No policy | 33% |
| Does not know whether a policy exists | 38% |
A prohibition does not remove the behavior. On this evidence it removes your knowledge of it, and knowing was the only instrument you had.
Why the prohibition is still in place
Almost nobody decided to ban AI tools on the merits. The ban arrived as a line in a legal or risk review in the first months after generative assistants became usable, when the honest answer to “what happens to our data” was that nobody knew. It was a sensible holding position.
What makes it persistent is that it has no owner. A sanction requires somebody to choose a vendor, sign a contract, configure a tenant and answer questions. A ban requires nobody to do anything, so it survives by default, and there is no meeting on anyone’s calendar where reversing it is the agenda item.
It also feels like it has already worked: no incidents in the log, no connector requests in the admin console, no AI line items on expense reports. A quiet system reads as a controlled system, and the study’s numbers suggest the quiet is the concealment.
The structural version of this problem, where the document and the behavior have no connection to each other, is the subject of our guide to safe AI use. The prohibition is the sharpest case of it.
What the cohort numbers prove, and what they do not
Treat the 67% figure carefully. These are self-reported cross-sections, not an experiment, so causation is not established. Organizations that ban may hold more sensitive data to begin with, and their employees may be more aware that what they did was prohibited and report it more readily. The banned group is also the smallest of the four, at 6% of respondents, which widens the error bars around its number.
What the data does establish is narrower and sufficient: there is no cohort in which prohibition is associated with low exposure. If bans worked as intended, the banned group would sit at the bottom of that table. It sits at the top. The data also shows that a policy alone barely moves the number, since guidance without education lands at 56%. Whatever closes this gap, it is not another document.
Sanctioning is four commitments, not one purchase
The security literature describes sanctioning as “provide an approved alternative,” as though the work ends at a purchase order. A sanction is a small product you ship to your own staff, and it has four parts that have to arrive together.
A named tool, not a category. “Use an approved AI tool” sanctions nothing. Name the product and the tier: the business or enterprise tier of one general assistant, bought under a contract with no training on your content by default, with single sign-on so the account belongs to the company rather than the person. One tool is easier to teach than three.
A data line people can recite. Not a classification schema. Four or five categories in the words your team already uses, plus one sentence on what to do when something falls between them. If an employee cannot repeat the rule from memory at 5pm on a Thursday, the rule is not operating.
An exception path faster than the workaround. A slow intake queue is the mechanism that produces shadow AI: the free tool is one browser tab away and the approval takes three weeks. Publish a named owner, a short request form, and a commitment to answer within five business days. Then meet it.
A capability commitment. Someone has to teach people to get more out of the sanctioned tool than they were getting out of the one they chose themselves. This is the enablement half of the work, and without it the other three parts are a tax.
Skip any one of the four and the sanction becomes a license nobody uses. The broader case for treating access and capability as different purchases is in why buying licenses is not the same as building capability.
The displacement is measurable
Sanctioning is not a hopeful theory. Netskope’s Cloud and Threat Report 2026 is built on cloud security telemetry from October 2024 to October 2025. Over that year, the share of generative AI users working in personal AI apps fell from 78% to 47%, while use of company-approved accounts rose from 25% to 62%. That is displacement at scale, and it happened as organizations deployed enterprise tiers. People did not have to be caught. They had to be given something better with their name on the company’s account.
The same telemetry carries the warning. Users switching back and forth between personal and corporate accounts rose from 4% to 9%. The switch is the signal that the sanctioned tool is losing on specific tasks, which makes it a curriculum problem rather than an enforcement problem, and the one metric here worth watching monthly.
The arithmetic for whoever signs
The budget conversation is short, because the numbers are not close.
On the cost side, a managed business seat is priced in the low tens of dollars per user per month. Anthropic publishes Claude Team at $20 per seat per month billed annually, with single sign-on and no model training on your content by default. Its Enterprise tier is published at the same seat rate with a 20-seat minimum and usage billed separately, adding SCIM, audit logs and retention controls. Check each vendor’s current page before you build a model; these prices moved more than once during 2026.
On the exposure side, IBM’s Cost of a Data Breach Report 2026 found shadow AI present at 43% of breached organizations, up from 20% the year before. Breaches involving shadow AI averaged $5.39 million against a global average of $4.99 million, and about one in five produced a reported fine.
Sanctioned seats for a 120-person company run around $30,000 a year at list, often less once you scope seats to the people who need them. The gap between a shadow-AI breach and an average one is about $400,000, roughly thirteen times that figure. Prohibition is also the option that puts the usage in accounts you cannot see.
We break down what a paid seat does and does not buy you in the comparison of enterprise accounts against consumer logins.
The part procurement cannot buy
Here is the failure mode that follows a successful purchase. The company buys seats, sends an announcement, and a quarter later most of the licenses are idle while the free tools keep running in parallel. Nobody is being defiant. The sanctioned tool is worse in their hands, because they learned the free one by trial and error over eighteen months and nobody taught them anything about the new one.
The study makes the scale of this gap explicit: within organizations that use AI, only about one in two employees in advanced economies report that their employer offers any training in responsible AI use. Prohibition and purchase are both cheaper than teaching, which is why both are more common.
A sanctioned tool wins when a person can do a real task in it faster and better than they could in the tool they picked themselves. That means their own documents, their own workflow, and the data rules applied to a live case rather than described in the abstract. Two hours of that does more for shadow AI than a quarter of policy circulation, because it changes which tool is the path of least resistance.
Two instruments belong alongside teaching rather than ahead of it. Technical inspection has a place once accounts are settled, and we sequence it in our guide to data-loss prevention for AI assistants. The enforcement side can be built without watching anyone, which we work through in enforcing an AI policy without surveillance.
Where a ban is the right answer
Prohibition is the wrong default. It is still the right call in narrow, nameable cases:
- A specific tool with disqualifying terms. A consumer product that claims a license to train on your inputs, with no business tier available, should be blocked by name. That is a tool decision, not a technology decision.
- A data class under a binding external obligation. Protected health information, cardholder data, or material covered by a client contract that forbids third-party processing. Prohibit the data path, not the assistant.
- A bounded legal window. A matter under litigation hold or an audit in progress justifies a dated suspension with a named end.
- Unvetted agentic connectors. A tool requesting mailbox or drive scopes through an OAuth grant is a different risk class from a chat window. Allowlist those explicitly.
Each of these is specific, defensible and reversible. “No AI tools” is none of the three.
Sixty days from prohibition to sanction
Days 1 to 10: find out what you have, with amnesty. Ask in writing, stating plainly that nobody will be penalized for an honest answer. Pair it with what is visible without inspecting anyone’s content: expense reports, single sign-on grants, browser extension approvals, and pending connector requests in your admin console.
Days 11 to 20: decide the tool and the data line. Pick one assistant and one tier. Write the four data categories. Get the exception path and its owner agreed by whoever would otherwise be the bottleneck.
Days 21 to 35: configure before you announce. Turn on single sign-on, set retention defaults, restrict personal-account logins to the sanctioned domain, and verify the no-training terms in the order form rather than the marketing page. A rollout announced before the tenant is configured teaches people the sanction is cosmetic.
Days 36 to 50: teach, on real work. Run applied sessions on the team’s own documents and recurring tasks. Lift the prohibition in the same communication that introduces the guidance, so the new rule arrives attached to new capability.
Days 51 to 60: set the three numbers you will watch. Share of AI activity in sanctioned accounts, count of distinct unsanctioned tools still in use, and time to answer an exception request. A rising switch rate means more teaching on the tasks where people switch, not a stricter rule.
Frequently asked questions
Does banning AI tools stop employees from using them?
No. In the KPMG and University of Melbourne global study, employees whose organization had banned generative AI reported uploading sensitive company information into public AI tools at 67%, compared with 33% where no policy existed. The study’s authors conclude that outright bans may be ineffective.
What does it mean to sanction an AI tool rather than ban it?
Naming a specific product and tier bought under a business contract, publishing a short data rule people can recite, running an exception path faster than the workaround, and teaching staff to use the sanctioned tool well. All four together. Any three of them produce licenses nobody uses.
How do we find shadow AI without monitoring software?
Ask with amnesty, then read what is already visible: expense reports, single sign-on grants, browser extension approvals, and pending connector requests in your admin console. A survey surfaces intent and unmet need that telemetry cannot see, and it costs nothing.
Will sanctioning a tool move usage off personal accounts?
The telemetry says yes. Netskope’s Cloud and Threat Report 2026 found personal AI app use among generative AI users fell from 78% to 47% in a year while company-approved account use rose from 25% to 62%. Watch the account-switching rate, which rose from 4% to 9% over the same period.
What does this cost compared with the risk?
A managed business seat runs in the low tens of dollars per user per month at published rates; verify each vendor’s current page. IBM’s Cost of a Data Breach Report 2026 puts breaches involving shadow AI at an average of $5.39 million against a $4.99 million global average, with one in five drawing a reported fine.
Who owns this when there is no Chief AI Officer?
Name two owners in the document: one for the tool list and data rule, usually operations or legal, and one for the controls, usually IT. At this company size the failure is vacancy rather than conflict, where each function assumes the other has it.
Should we keep the ban in place while we roll out a sanctioned tool?
Lift it in the same communication that introduces the replacement. A prohibition that outlives its replacement teaches people that neither rule is real, and it keeps the concealment behavior running during the period you most need honest reporting.
What if we cannot buy anything this quarter?
Three of the four commitments cost nothing. Write the data rule, publish the exception path with a named owner and a response time, and run one applied session on real work. Then move the usage you cannot fund onto the most defensible free tier available and document that as an interim position with a date.
Key takeaways
- Employees at organizations that banned generative AI reported putting sensitive company information into public AI tools at 67%, against 33% where no policy existed. The ban correlates with the behavior it forbids.
- Seventy percent of employees who use AI at work reach it through free public tools; only 42% use anything their employer provides.
- A sanction is four things shipped together: one named tool and tier, a data rule people can recite, an exception path faster than the workaround, and someone teaching the tool.
- Displacement is measured, not hoped for. Personal AI app use among generative AI users fell from 78% to 47% in a year while approved-account use rose from 25% to 62%.
- Keep prohibition for named cases: a tool with disqualifying terms, a data class under external obligation, a dated legal hold, and unvetted agentic connectors.
If your AI policy is still a prohibition and you want a read on what it would take to replace it with something your team will use, book a free AI-readiness call. We will talk through where your people are working today and what a hands-on session on your own documents would need to teach for the sanctioned tool to win.
Dirk Jan van Veen, PhD