A manager we’ll call Dan has a report who’s been slipping for two quarters. The deadlines keep moving. The work comes back thin. Dan has a redirect conversation scheduled for Thursday, and he’s dreading it. So on Wednesday night he opens ChatGPT, pastes in a paragraph describing the situation, and types: “How do I tell someone their performance isn’t cutting it without them shutting down?”
He gets a clean, structured answer in about eight seconds. It’s calmer and more organized than anything he’d have improvised on his own. He tightens his notes, closes the tab, and goes to bed feeling more prepared than he has all week.
Notice what just happened, and what didn’t. No data breach. No records exposed. No compliance line crossed in any way Dan would recognize. A manager got coached on one of the most consequential moments in his job, and that coaching took place somewhere no one in HR will ever see. The security team has a name for this. They call it Shadow AI. HR should recognize it as something else first.
Someone else is already coaching your managers
When a manager asks a model how to structure feedback, defuse a conflict, or word a review, that’s a leadership-development interaction. It’s informal and unsupervised, but it’s real. Someone with authority over other people’s work is getting advice on how to use that authority, and they’re acting on it. Strip away the interface and it’s the same thing an L&D team spends budget trying to provide: guidance at the moment a manager actually needs it.
This isn’t rare behavior at the margins. According to the Microsoft and LinkedIn 2024 Work Trend Index, 78% of AI users bring their own AI tools to work, frequently without IT’s knowledge, a pattern the report labels “bring your own AI.” That’s not a fringe group experimenting. That’s the default way people have started using these tools, and they’re doing it faster than any policy can keep up.
Managers aren’t the light users in this picture. They’re the heavy ones. Reporting from TrustedTech via Help Net Security found that 65% of decision-makers use shadow AI, compared with 31% of non-managers. Read that gap slowly. The people using unsanctioned AI the most are also the people making decisions about other people, about who gets a hard conversation, who gets promoted, who gets managed out. The coaching they’re absorbing on the side is feeding directly into those calls.
If you want the fuller version of this argument, we’ve written before about what AI coaching for managers actually is and where it fits, and about the honest limits of the tools in our piece on AI versus human coaching. The point here is narrower and more uncomfortable. Whatever you think AI coaching should be, your managers have already started doing it. They just did it without you.
What managers actually turn to AI for in people work
It helps to be specific about what this looks like day to day. Not the abstract “employees use AI,” but the concrete moments managers reach for it. Four come up again and again.
Structuring difficult feedback and performance reviews
This is the big one. Managers use AI to find the words for feedback they’re nervous to deliver, to soften a review without gutting the message, and increasingly to make the language feel defensible if it’s ever questioned later. The instinct is understandable. Most managers were never taught how to give constructive feedback well, and a model that produces tidy, professional phrasing feels like a lifeline. The trouble is that tidy phrasing and honest, useful feedback aren’t the same thing, and a model optimizing for the former will happily sand the latter down to nothing.
Rehearsing conflict conversations before they happen
The second pattern is rehearsal. A manager knows a conversation is going to be tense, so they run it past the model first. What will they say back? How do I respond if they get defensive? It’s a genuinely reasonable use, close to what a good coach or mentor would offer. It’s also exactly the kind of preparation most managers currently do alone, at night, with no one watching. Our library of difficult conversation examples exists precisely because this need is so common and so poorly served inside most companies.
Getting a second opinion on a call they’re unsure about
Managers also use AI as a sounding board for judgment calls. Should I put this person on a formal plan or give it another month? Is this promotion case strong enough to bring forward? How do I reshuffle the team without wrecking morale? These are high-stakes decisions, and the manager is asking a model with no knowledge of the person, the history, or the organization to weigh in on them.
Managing up or across without a peer to sound-board with
The last one is quieter but common, especially among newer or more isolated managers. They don’t have a trusted peer to think out loud with, so the model becomes the peer. How do I raise this with my own boss? How do I push back on a deadline without looking difficult? We’ve written about the skill of managing up because it’s rarely taught, and when there’s no human to practice it with, an always-available model fills the vacuum.
None of these four uses is reckless. Every one of them points at a real gap in support that managers are trying to close on their own. That’s the part worth sitting with. This behavior isn’t a discipline problem. It’s a demand signal.
Why this is a leadership-development problem before it’s a security one
Here’s the part the current conversation about shadow AI keeps missing: the real risk is that the coaching managers get from these tools is invisible, inconsistent, and disconnected from anything the organization has decided it wants its managers to do, whether or not any data ever leaks.
Picture the same situation running through two managers. Both have a report struggling badly enough that a performance plan is on the table. Dan describes his to a model in three sentences with no context on company policy, no read on the person’s history, no sense of the values his organization says it leads by. He gets advice that sounds authoritative and is partly wrong, because the model is filling in the missing context with plausible guesses. Now take a second manager, Rachel, facing a nearly identical situation. Rachel prompts her model carefully and gets advice that’s perfectly reasonable, and identical to what three other managers got that same week, because that’s what the model outputs by default when you ask it a common question.
Neither of these is a security incident. No data leaked. But look at the outcome. One manager acted on advice that was confidently wrong for her specific person, and the other rolled out the same generic playbook as everyone else, flattening exactly the judgment that’s supposed to distinguish a good manager from an average one. Both are leadership-development failures. Neither would ever show up in a data-loss report.
And this compounds, because managers set the tone for their teams. Gallup’s 2026 research found that employees whose manager supports AI use are far more likely to use it themselves, 78% versus 44%. Sit that number next to the earlier one. Managers are already the heaviest shadow-AI users, and their teams take direct behavioral cues from them. So a manager’s own unmanaged AI habits aren’t a private matter. They’re quietly setting the norm for how an entire team uses these tools, good habits and bad ones alike.
The scale of this is almost certainly larger than HR thinks. The McKinsey State of AI 2025 report found that leaders consistently underestimate how much their own people are already using generative AI, by roughly three times. Apply that correction to your own manager population. Whatever fraction you’d guess is using AI for people decisions right now, the honest number is probably a lot higher.
Why banning it fails, and ignoring it is worse
Some organizations have responded by drawing a line: no using AI for performance decisions, feedback, or reviews. On paper it looks responsible. In practice it mostly changes what you can see, not what managers do.
The Wakefield and PagerDuty 2026 Shadow AI Survey puts numbers on this. Two-thirds of employees, 66%, used AI even though they believed it violated policy. A third, 33%, deliberately hid that use to avoid a manager’s scrutiny. And 39% said they wouldn’t disclose their AI use even if asked directly. Read that as a visibility problem, not a character flaw. A ban doesn’t remove the behavior. It converts a visible practice into a hidden one, and trades the chance to shape it for the comfort of pretending it stopped.
Ignoring it isn’t the safe middle ground either. Doing nothing is a decision too, and this is what it decides. By default, you’ve outsourced a piece of your leadership-development function to whichever model each manager happens to prefer, with no visibility into what advice is being given, whether it’s any good, or whether managers are acting on it. That’s a real program running inside your company right now. It’s just one nobody designed, nobody owns, and nobody can inspect.
So the real choice was never AI or no AI. That decision already got made, quietly, by your managers. The choice in front of HR and L&D is managed or unmanaged.
What responsible adoption actually requires HR and L&D to decide
Managed doesn’t mean a fresh stack of surveillance tooling, and it doesn’t mean a policy PDF nobody reads. It means making a few deliberate decisions that most organizations have been avoiding. This isn’t the full framework, and the rest of this series works through these questions in detail. But three of them are worth naming now.
What “good” looks like, and whether AI is meeting that bar
First, decide what good coaching advice for your managers actually sounds like. What do you want a manager to do when a report is underperforming, when a conflict flares, when a promotion case is borderline? Until you’ve named that, you can’t evaluate whether the AI advice managers are already getting clears the bar or falls short of it. Right now, most companies are letting a general-purpose model define “good” by default, without ever having decided what good means to them.
Where AI-assisted people-decisions should be visible, and where privacy should hold
Second, draw the line between learning and deciding. A manager privately working on their own skills should keep that privacy, the same way you wouldn’t demand a transcript of their coaching sessions. But a decision that affects someone’s role, pay, or standing is a different category. HR has a legitimate interest in how those calls get made and what informed them. The distinction between self-driven development and decisions that land on another person is where thoughtful visibility belongs.
Whether managers get any guidance at all
Third, and simplest: are your managers getting any guidance on how to use AI for people conversations, or are they left to work it out alone at 11pm? Most organizations have offered nothing here, which is precisely why the shadow version filled the gap. Guidance doesn’t have to be heavy. But its total absence is a choice, and it’s the choice that produced the situation you’re now trying to manage.
Working through these is the practical core of building a real business case for leadership development in an environment where AI is already in the room whether you invited it or not.
Where this leaves HR and L&D
The organizations that get ahead of this are the ones treating shadow AI among managers for what it actually is: the largest, least-supervised leadership-development program most companies are currently running, and one nobody chose to run. Banning it only pushes the behavior further out of sight. Ignoring it hands the job of setting the standard to whatever model your managers happen to open. Treating it as a coaching question, not a security incident, is what actually gives HR a say in the outcome.
That reframe is the whole point. Your managers didn’t wait for a rollout plan. They’re already being coached. The only open question is by whom, to what standard, and whether anyone in HR has a say in the answer.
Give your managers a better default
If managers are going to be coached by AI either way, the responsible response is to give them AI coaching that was actually built for the work, so the tool they reach for at 11pm is one you’d stand behind. That’s the real distance between a general chatbot and a coaching platform, and it’s worth being specific about.
It coaches, instead of answering
A general model hands a manager an answer. Purpose-built AI coaching works the way a good human coach does: it asks questions, surfaces the blind spot the manager can’t see on their own, and asks for a concrete commitment before the conversation ends. The goal is a behavior that changes. We’ve written about where that line sits in AI versus human coaching.
It remembers
A chatbot forgets the moment the tab closes. A coaching platform holds the thread across months, so it can tie the delegation habit a manager set as a goal in one session to the 1:1 they’re preparing for weeks later. Memory is what turns a good answer into coaching.
It measures
ChatGPT can’t tell you whether a manager actually got better. A coaching platform assesses skills at the start and at intervals, calibrated with 360 feedback so a score reflects how a manager’s team experiences them, not just how the manager rates themselves. That’s how you get from “our managers are using AI” to a number you can defend: an average 26% improvement in coached skills over 12 weeks.
It stays visible to HR without exposing the person
This is the part shadow AI can never offer. Self-directed coaching stays private to the manager. For coaching plans the organization assigns, HR sees session summaries and engagement signals, not transcripts, and conversation content isn’t used to train models. Leaders get the aggregate skill trends they need to run a program while managers keep the candor that makes coaching worth doing. A public chatbot gives you neither the privacy guarantee nor the visibility.
Put those together, coaching that changes behavior, memory that compounds, measurement HR can report on, and privacy that keeps managers honest, and you have what a manager pasting a performance plan into a public model at 11pm simply doesn’t get.
This is the problem Risely’s AI coach, Merlin, was built to solve: coaching managers can reach for in the exact moments they’re already opening ChatGPT for, available inside Slack and Microsoft Teams, with the visibility HR can trust and the privacy managers need. For the point-by-point version of the tradeoff, we’ve laid it out in Risely versus ChatGPT for coaching.
Your managers have already chosen AI coaching. The only decision left is whether they get a version built for the job. See what a sanctioned alternative looks like with Merlin.