Tomorrow morning a manager has a performance conversation they have been putting off. They mostly know what needs to be said. The harder part is what happens after the first sentence: the pushback, the silence, the question they did not prepare for. There is now an obvious place to go before that meeting, and it takes about ten seconds to open. A public AI tool.
A difficult conversation, in the sense used here, is a work conversation where at least one person stands to lose something and where the path it takes cannot be predicted in advance: a performance review, critical feedback, a pay discussion, a termination. The need to prepare for those is not new. What is new is where the preparation now starts.
We don't know how universal that behavior is. A new survey from The Predictive Index does give us a timely signal that it is no longer an edge case.
What did the new survey actually find?
Predictive Index is an HR technology company working in behavioral assessment and people analytics. Its AI and People Leadership Report was fielded in July 2026 and published on 18 August, across two samples: 399 managers and 208 CEOs and business leaders.
The headline number is 72%. That share of managers said public AI tools are useful when preparing for people-related conversations. The more concrete behavior sits at 44%, the share who said they had entered employee names or performance details into those tools.
Another figure matters just as much. Only 45% said their organization had a formal written policy governing managers' use of AI in people matters, while 68% of leaders said they were worried about employee information going into public tools.
The survey's own headline is elsewhere. 74% of leaders believe their managers are "very confident" handling tough people conversations on their own. Only 20% of managers say they would rather prepare for one alone, without a framework, a coaching guide or input from HR.
And then the number that explains the most while making the least noise: 42% of managers said they know what they want to say but struggle with how to say it.
None of this is peer-reviewed, and it shouldn't be read as if it were. Predictive Index is a company with a commercial position in this space, and the public release gives sample size and broad composition without the questionnaire, the weighting or the margin of error. The methodology paragraph says "two national samples" without naming the country. So the narrow claim is the useful one: in this July 2026 sample, managers treated public AI as a serious preparation tool, a substantial share were already carrying real employee context into it, and written policy had not kept pace.
The more interesting number is 42%
72% tells us the tool looks useful. 42% tells us where managers are actually stuck. For most of them the substance of the message is already there. It's the delivery that gives way.
Putting 74% next to 20% is instructive too, as long as it's read carefully, because the two figures measure different things. One is what leaders believe about their managers; the other is what managers say they prefer. Even so, the distance between them says something. Leaders think their managers can carry these conversations alone. The overwhelming majority of managers would rather not, and some of them are closing that gap with whatever tool is nearest.
Set 44% beside 45% and a second gap appears. Employee context is moving into public tools, and in most of the same organizations nothing written says where that use should start or stop.
The question is no longer whether managers will use AI. The more practical one: which part of the preparation belongs in which environment, with what employee data, and under whose rules?
A good draft is not a good conversation
A public chatbot can suggest a better opening line. It can take the heat out of a draft, list the objections worth expecting, even play the other side for a few turns. All of that is useful.
It still doesn't make an ad hoc chat the same thing as a repeatable practice environment that an organization can observe against consistent criteria and review afterwards.
The real performance review begins after the prepared line. The employee pushes back. Goes quiet. Gets angry. Puts an unexpected fact on the table. The script ends there, and behavior begins. One person can prepare the first sentence. The second one is settled by two.
We've written about why these conversations need practice rather than wording, with the research behind it, in difficult conversations at scale.
Both sides of the table are preparing
The survey looks at managers, but the conversation takes two. Someone is preparing on the other side as well: the one about to ask for a raise, to push back on a rating, to carry a problem upward, to say no to a deadline.
Control of a prepared script ends the moment the other person starts talking. What neither side actually needs is a better opening line. What they need is to have already been through what comes after it.
Role-play is the method, not the outcome
The whole point of a rehearsal is that you can still get it wrong before the performance starts. For difficult conversations that room has been missing.
↳ SOURCEA few years ago, an AI character that answered convincingly was enough to carry a demo on its own. That capability is becoming standard. The durable question for an organization isn't whether the character talks, it's what the conversation makes visible.
So role-play is the method here, not the product. Our own reading is that the useful information is never the line someone rehearsed. It is what they do once that line stops working. Raise their voice, retreat, change the subject, or take the objection and rebuild the conversation around it.
That changes the question worth asking when evaluating a tool. Does the character actually resist, or does it just respond? We collected that one and three others in a separate piece on evaluating AI role-play.
From one session to team readiness
One session belongs to one person. When the same kind of situation repeats across dozens of people, an organization can start asking different questions.
Which conversations keep proving difficult? Which behaviors deteriorate as pressure rises? Which teams need more practice in the same situation? Does the pattern soften after training?
Those are management questions, not a feature list, and they point at a different unit of value than "17 people completed the course." The organization begins to get an early view of how ready it is for a situation it can't safely observe in real life.
An early view, not ground truth. We've gone through what a single score can and cannot become at team level in beyond the score screen.
Employee data is part of the design problem
This is why the 44% finding isn't a side detail. Realistic preparation needs context, and context usually means real information about a real employee. Moved into an uncontrolled tool, that same context becomes a separate risk.
So it isn't enough for an enterprise answer to give better advice. What data can be used, who can see it, how long it stays and which provider it reaches are part of the design too.
What we take from this at EVRE
The most interesting thing in the survey isn't that managers are using AI. It's the moment they reach for it: immediately before a high-stakes conversation with a real person.
That's why we don't build EVRE as a tool that answers "what should I say?" The user goes into the conversation instead. The other side answers, the tension moves with how the conversation is going, and the feedback at the end is tied to what was actually said. One session gives one person feedback. Repeated sessions produce a signal about readiness at team level.
Reaching for AI before a difficult conversation may be a new managerial habit. The need to prepare isn't new at all. What's new is that the preparation can now begin in a public tool, in seconds, with no decision made about it anywhere.
The real choice for organizations won't be between banning AI and allowing it. It will be designing where preparation happens, with what data, and how closely that preparation should resemble the real conversation.
Generating the draft of a difficult conversation is one problem. Staying effective after the other person answers is another.
