The Conversations Your Managers Are Already Having With AI

Somewhere in your organization this week, a manager will open ChatGPT or Claude and type a version of the following: "I have to give difficult feedback to someone on my team and I don't know how to start. Here's the situation."

This is happening. It is not a forecast. In April 2026, Anthropic published How people ask Claude for personal guidance, an analysis of one million conversations on its consumer platform. It found that 26% of personal-guidance conversations were about professional and career questions, second only to health and wellness. People are already turning to AI to think through the hardest parts of their work lives. The conversations they are trying to have with AI are the same conversations a coach would have with them, if they had a coach.

Most of them do not. Traditional executive coaching, at $300 to $500 per hour, has historically been available to a small fraction of any organization, typically the senior executive layer. The middle management layer, where the conversations that drive retention and performance actually happen, has been left to figure it out alone.

Now they are figuring it out with AI. The question for HR leaders is what they are doing to make sure those conversations actually help the people having them.

What the Anthropic research found

Base-model Claude exhibited sycophantic behavior — excessive agreement, validation of one-sided framings, and reluctance to push back — in 9% of guidance conversations overall, and 25% of relationship conversations, according to the same Anthropic study. When users pushed back on the model's read, the rate roughly doubled, from 9% to 18%.

Under pressure, generalist AI becomes more agreeable, not less.

For the kinds of conversations managers are bringing to AI, this is the wrong direction. The performance issue they have been avoiding. The hire they are about to make for the wrong reasons. The team conflict they are mishandling. The career decision they have already half-made and want validation for. These are precisely the conversations where a manager is least likely to want a clear read, and most likely to push back when they get one. A guidance tool that softens its position under pressure is not coaching them. It is confirming whatever they already wanted to do.

What separates coaching from a chatbot

The honest reading of the Anthropic data is not that AI guidance is dangerous. It is that the quality of AI guidance depends entirely on what is built around the model. The base capability is powerful. What turns it into coaching is everything that surrounds it.

Three things matter, and none of them are user-configurable in a generalist tool:

Methodology. Real coaching operates from named principles. Candor as a professional obligation: a coach who sees a flaw in reasoning and stays quiet to preserve comfort is failing the user. Earned directiveness: perspective offered after the coach has actually understood the situation, not before. The ability to read what kind of conversation is happening (exploratory, directive, or decisive) and shift mode accordingly. These are not personality traits of a model. They are design decisions, specified at the level of the framework that governs every conversation. A manager typing "be more direct with me" into ChatGPT does not produce coaching behavior. The framework has to be built in, and someone has to have thought hard about what it says.

Continuity that compounds. A coach remembers what the manager committed to last week and asks how it went. A coach references the framework they built together two months ago. A coach sees patterns across sessions that the manager cannot see in any single one. The work gets more valuable over time, not less, because every session builds on the one before it. Generalist AI tools start fresh every conversation. Continuity is not a feature a user can turn on. It has to be built into the system.

Organizational context. A coach inside an organization understands the values, the leadership principles, and the strategic priorities that shape what a good answer looks like. A generalist AI tool knows none of this, and cannot. Organizational context is something the platform has to be built to carry. It is not something the manager can paste into a prompt.

The institutional question

For a CHRO, the question is not whether managers should be allowed to use AI for guidance. They already are. The question is whether the organization has chosen to provide structured coaching infrastructure built for that use, or whether it has effectively defaulted to whatever consumer tool each manager picks up on their own.

The first is a development strategy. The second is the absence of one.

The shift that has happened, quietly, over the last two years, is that AI has made coaching at scale economically possible for the first time. The traditional barrier was cost. A coaching engagement at $300 per hour, extended to a full middle management layer, was a budget line no mid-size organization could absorb. That math has changed. What has not changed is the fact that the conversations that drive retention, engagement, and team performance still need to happen with something more than a generalist chatbot can provide. Actual coaching happens when the coach listens, internalizes with context, asks powerful questions, offers principle-based feedback and guidance, and is consistent and informed over time.

Anthropic's research, published by the company that makes one of the largest generalist AI models in the world, is itself the clearest statement of why the distinction matters. The base capability is powerful. The coaching is what gets built on top of it.

The managers in your organization are already having these conversations. The only open question is whether they are having them in an environment built for the work, or in a tool that was built for something else.

If your organization is ready to look at what coaching infrastructure for the middle management layer looks like, we'd welcome the conversation. Let's Chat Coach offers free trials and pilots.