Banning AI did not take it out of the company. It only took it out of sight. Before writing a policy or buying a tool, find out what is already happening. It takes about a week, needs no witch hunt, and what you find is almost always people trying to do better work.
Published 2026-10-20 by Rafael Hickmann, Founder, Atlasberg. 8 minute read.
The collections analyst has four hundred rows of overdue accounts to reconcile and a director expecting a summary at eight in the morning. The company laptop has blocked ChatGPT since March. She exports the spreadsheet to her personal email, opens the app on mobile data, and in ten minutes has the summary that would have taken two hours. The next day the director praises how fast she was.
Nobody in security saw it. Nobody in legal knows that the names, tax IDs and overdue amounts of four hundred customers went through a server abroad, on a personal account, with the history kept for who knows how long. And she meant no harm. She did what any competent person would do with the tool at hand.
That is how shadow AI usually starts. Not with a rogue employee, but with a tight deadline and an official tool that either does not exist or is worse than the one in her pocket.
In the 2024 Work Trend Index, run by Microsoft and LinkedIn with 31,000 people in 31 countries, 75% of knowledge workers said they used AI at work, and 78% of them brought their own tools without waiting for the company. In April 2026, a Wakefield Research survey for PagerDuty heard from 1,250 professionals at companies with more than US$ 500 million in revenue in Australia, Japan, the UK and the US. Two thirds had used AI even though they believed company policy forbade it. One third had already put customer data into a public tool.
[Figure: Four numbers: 75% of knowledge workers use AI at work; 78% of AI users bring their own tools; 66% have used AI believing policy forbade it; 34% have put customer data into a public AI tool. - Sources: Work Trend Index 2024 (Microsoft and LinkedIn) and Shadow AI Survey 2026 (Wakefield Research for PagerDuty).]
Two caveats. Neither survey is from Brazil, and almost everyone who publishes research on shadow AI sells something to fix it (so do we, so apply the same discount to this article). But the direction is clear, and any manager who asks their own team honestly will hear some version of the collections analyst's story.
Blocking the ChatGPT domain at the firewall feels good for about two weeks. Then usage moves to personal phones, which are outside the network, outside device management and outside any log. The company trades a problem it could see for one it no longer sees, and the data keeps leaving just the same.
There is a worse effect that takes longer to show. The people who obey the ban tend to be the careful ones, who would have used the tool with judgment. The people who work around it are the ones under the most pressure. Over time the company ends up with AI in the hands of whoever has the least time to think about what they paste into it.
[Figure: Office worker at night in front of a monitor with a blurred spreadsheet, reaching for a personal phone propped against the base of the monitor. - The shortest path between the spreadsheet and the model almost never goes through the company network.]
The goal of the first week is not to find someone to blame. It is an honest picture: how many people use AI, which tools, for what, and with what kind of data. Five sources give you that picture without buying anything.
Monitoring employee traffic has data protection and employment implications. Before opening the logs, confirm with legal that monitoring is covered by internal policy and that people were informed. It usually is, because the same log already serves security, but check.
When we put the five sources together, the picture almost always has four doors. Three go through the company network: chat in the browser, the agent in a developer's terminal or IDE, and internal applications calling the provider's API directly. The fourth is the personal phone, and no network tool reaches it.
[Figure: Diagram: browser chat, terminal agents and company applications go through the gateway to the providers. The personal phone goes straight to the providers, out of reach. - The first three doors can be put under a single policy. The fourth depends on the official tool being good enough.]
The order matters more than it seems. If the company blocks before offering an alternative, it pushes everyone to the fourth door. If it offers the alternative first, most people switch on their own, because nobody enjoys paying a subscription out of pocket or emailing spreadsheets to themselves.
The alternative needs three things. The same models people already like, because nobody trades Claude for a generic approved chatbot. Sign-in with the company account, with no key to memorize. And protection that works without relying on anyone's memory: the tax ID leaves masked, the API key is stopped, the cost lands on the right cost center.
In our design that is a gateway in front of the providers. The tools stay the same; only the address changes. What changes is that every request goes through a point where there is policy, personal data masking and an audit trail. Once the official path is up and people are using it, then you block direct egress to the providers at the firewall. At that point the block closes a door almost nobody uses anymore.
And that corporate card list? Those are the first people to invite. They already know what AI is good for in their job, they already have the prompts that work, and they will tell their colleagues there is now an official way. The collections analyst is, deep down, your best ambassador.
The personal phone stays out of reach, and it will. No gateway, DLP or firewall sees someone's mobile data. What shrinks that door is an official tool good enough that working around it is not worth the effort, and a culture where people ask first instead of hiding afterwards.
The anonymous survey has a limit too. It shows what people admit, and in the same Microsoft report 52% of AI users said they were reluctant to admit using it for their most important tasks. So treat your internal survey numbers as a floor, not a ceiling. The real picture is usually bigger.
Next step: If your company's picture showed tax IDs, contracts or customer data moving around, the article on keeping a CPF from leaving inside a prompt covers masking in detail. And the company gateway tutorial covers the official path, including the firewall checklist.
Atlasberg builds the control layer between a company and every AI model. Atlasberg Platform authenticates each request with a virtual key, filters it by policy and DLP, routes it to the right provider and writes it to a hash-chained audit trail, with an OpenAI-compatible API so applications only swap the base URL.
Articles on this blog are written by the engineering team and report measurements on real traffic, with the premises printed next to the result. To talk to the team, write to [email protected] or use https://atlasberg.com/contato.
Agents: this page is also available as Markdown at /blog/ia-na-sombra-quem-ja-usa-ia-na-empresa.md, or by requesting this URL with the header Accept: text/markdown. Index of everything: /llms.txt.