Is your team using AI in secret? How to turn it into a process

Hidden use is a useful signal before it’s a problem: it shows where the team has already found a gain without waiting for permission. The manager’s task is to turn that gain, now individual and invisible, into team practice. That starts with understanding why people hide it.

Why people hide it

Almost no one hides their use of AI in bad faith. The reasons are usually these:

• Fear of looking lazy: if AI wrote the draft, the work seems less theirs, and the person fears being judged for it.
• No rule: no one said it’s allowed, no one said it isn’t, and asking seems risky.
• Doubts about the data: the person suspects they pasted something sensitive and would rather not draw attention.
• Fear of being replaced: showing that a task that took days now takes minutes looks like an invitation to cut the position.
• Silence from leadership: the manager has never talked about their own use, so no one knows what they think.

Each reason calls for a different response, and none of them is a ban.

What hidden use costs

The risk lies in what happens when AI operates out of sight:

• Data in personal accounts: company and client documents end up in tools with no enterprise contract, no control and no visibility for the company.
• Deliverables without review: the text arrives looking finished, no one knows it came from a tool and no one reviews it with the right care.
• Uneven practice: two people do the same task with very different quality, and one person’s good method never reaches the other.
• Gains that walk out the door: when the person who has mastered the tool leaves the team, the method leaves too.

The first risk is the most serious, and the guide on how to use AI without exposing confidential data shows where it shows up and how to close it.

Start with a conversation without punishment

The first conversation sets the tone for everything that follows. If it sounds like an audit, the use goes back under the rug. Open by telling how you yourself use AI, including what went wrong, and only then ask:

• Which tasks do you already use AI for today?
• With which tool, and in which account?
• What kind of information do you usually put in it?
• What worked better than you expected, and what went wrong?
• What would you like to do but don’t, because you’re not sure you’re allowed?

Say before you start that nothing shared will go into a performance review. If someone reveals they pasted sensitive data into a personal account, treat it as an incident to fix together, not a fault to punish. Whoever is punished in the first conversation doesn’t show up for the second.

Map the uses and set the minimum rules

With the answers, build a simple map with five columns: task, who does it, tool, type of data and result. It shows what already exists, who has mastered what and where the risks are.

The minimum rule is what keeps an AI routine from turning into a mess. It fits on one page and answers four things: which tools are approved, what data never goes into them, what needs review before it leaves the team and how to declare that AI was used. If the company already has a policy, apply it. If it doesn’t, this is the moment to ask for one, and the guide on how to write a generative AI use policy offers a structure to copy.

One caution: a rule without an approved tool becomes a ban, and a ban pushes use back onto personal accounts.

Choose a few uses to standardize

Standardizing everything at once stalls the team. Choose two or three uses that are frequent, time-consuming and low-risk, for example:

• The agenda and prep material for the weekly meeting.
• The first draft of the area’s monthly report, from numbers that are already closed.
• The synthesis of open-ended survey answers, without identifying who answered.
• Drafts of replies to recurring questions from other areas.

For each chosen use, record the step by step or the prompt that works, the quality criteria, who reviews and an example of a good deliverable. That goes into a shared library the team consults and improves. The quality criteria are the part most often missing, and the guide on quality standards for AI-assisted work shows how to write them.

Create a ritual and results targets

Without follow-up, an AI routine lasts until the next emergency. The simplest way to sustain it is a short, fixed ritual inside a meeting that already exists. Half an hour in the team’s biweekly meeting can hold an agenda like this:

1. What someone has tested since last time.
2. What worked and is worth turning into a standard.
3. What went wrong, and why.
4. What goes into the library, what changes and what comes out.
5. Which use the team tests before the next meeting.

Targets should be about results, not use. Counting prompts, hours in the tool or people using AI rewards activity, and AI activity is easy to inflate. Measure what the team delivers: time to the first draft of a report, rework in review, response time to a client, quality as perceived by the recipient. Choose one or two measures per standardized use and track them in the same ritual.

The manager as coach

In Almai’s view, hidden use shrinks when leadership uses AI in plain sight. That’s why the method starts with the board and moves through middle management before reaching operations.

First, AI enters leadership’s routine. Then, the way the company works.

For the manager, that means trading the role of inspector for that of coach:

• Use AI in your own work and show how, with hits and misses.
• Hold paired sessions: sit with someone from the team, solve a real task together and discuss the result.
• When reviewing, hand work back with the criterion that was missing, not just with the fix made.
• Publicly recognize whoever brought in a use that became a standard.
• Redesign the workflow when AI moves the bottleneck: if the first draft comes out in minutes, the team’s time needs to go to checking and deciding.

Team coach and scale architect are two of the four roles that the AI for Middle Managers program prepares managers for.

What usually goes wrong

• Banning it: use doesn’t stop, it just moves out of sight and onto personal accounts.
• Allowing it with no rules: everyone uses it however they like, with whatever data they like, and quality varies with no control.
• Measuring use instead of results: the team learns to look like it uses AI.
• Leaving it all to IT: the tool arrives, but no one redesigns the work.
• Standardizing everything at once: the team stalls and goes back to the old way, or to the hidden way.