What to Do About the AI Your Team Already Uses

A company's security team cross-referenced its own network logs and found that a good part of the staff was already uploading internal documents to an AI nobody had approved. The response was to block access that same afternoon. Two weeks later the traffic came back. Nobody could see it anymore.

Carlos Andrés Ramírez · 2026-09-14

A company's security team cross-referenced its own network logs and found that a good part of the staff was already uploading internal documents to an AI nobody had approved. The response was to block access that same afternoon. Two weeks later the traffic came back. Nobody could see it anymore.

It happens in almost every company that actually looks. Someone audits outbound traffic, or reviews the corporate card statements, or just asks in a meeting who has tried this, and the same surprise shows up: people are already using AI, on personal accounts, on the free tier, uploading whatever the work requires to hit the deadline. Nobody decided this in a committee. Every person decided it alone, on some ordinary Tuesday, because the approved tool took weeks to clear and the deadline did not wait weeks.

The first instinct is almost always to shut it down. Block the domain, add the AI to the proxy blacklist, send a reminder about acceptable use. It feels like action. And it is, except it solves the wrong problem. The risk was never that the AI existed. It was that nobody knew what data was leaving through it. Blocking access kills the one signal the company had to measure that risk, right when it could finally see it.

The symptom

What does a company do when it finds out its people are already using AI without permission (shadow AI)?

The reflex answer almost everyone gives is ban first, ask later. Legal sends the notice, IT closes the access, the committee exhales because it did something visible. But the use does not disappear. It moves. Someone who already found that AI saves hours of work does not go back to the manual method because an email arrived. They switch devices, use a personal phone, log in from a network the company does not watch. The company gets a memo and loses the visibility it had by accident.

• The first sign of unauthorized AI use comes from an expense or traffic audit, almost never from someone reporting it.
• The acceptable use policy has existed for a while, but nobody updated it since AI tools changed this fast.
• The team using unauthorized AI the most is usually the one under the most delivery pressure, not the one that understands the risk least.
• The company's response gets decided in a same-day crisis meeting, before anyone measured what data went out or where.
• Blocking the domain on the corporate network does not reduce use. It moves it to a personal phone, where the company sees nothing at all.

The problem underneath

The risk is not that your people use AI. It is banning it before you look.

When a company discovers unauthorized AI use, it has, without realizing it, information it did not have the week before: real demand exists, this is the kind of task people solve with it, and these are the people doing it. That information is worth more than what the company had a month ago, when the use was already happening and nobody had looked. Banning it immediately throws that information away before it gets used for anything.

And there is a second loss, costlier than the first. Every time an employee finds a faster way to work and the company takes it away without offering an equally fast alternative, they learn a lesson that sticks: solving the problem on your own works better than asking permission. The next tool they find, they will not report. They will hide it better.

Banning the AI your people already use does not remove the risk. It removes it from view, the same day you finally had it in front of you.

BECOME

The framework

What should happen in the first weeks after discovering unauthorized use?

The five pieces below get worked in order, not all at once, and in the first weeks after the first signal shows up, not months later as part of a fully built governance project.

Real scope

What is being used, with what data, and across how many teams, before deciding anything. Without this map, whatever policy gets written ends up solving a problem nobody measured.

Amnesty window

A short, clearly communicated period in which declaring existing use carries no disciplinary consequence. Without it, nobody comes forward and the map from the first step stays incomplete.

Data red line

Which categories of information are off limits with no exception, such as client data, contracts, or regulated information, while everything else waits for a calmer decision. Not all use carries the same weight, and treating it as if it did is what makes the ban feel arbitrary.

Sanctioned fast path

An approved alternative that takes as long to clear as it takes an employee to open a new tab, not several weeks. If the official route stays the slowest one, unauthorized use comes back within a month.

Risk owner

One named person with the authority to decide, case by case, what gets tolerated while the permanent policy gets built. Without that owner, every manager improvises their own rule, and the company ends up back where it started.

None of the five requires turning the company into an AI police force, or treating whoever already uses it as if they did something wrong. It requires the opposite: recognizing that this person found a real problem and a faster answer than the one the company offered, and building the version of that answer the company can actually sustain.

Before writing a single line of policy, ask three different teams what AI tool they already use today, without anyone having approved it. If the answer comes back fast and straight, there is still a window to sort this out without losing the trust of whoever is about to tell you. If the answer is silence, that window already closed, and it has to be earned back.

Let's get a handle on the AI you already have