Why YouMight Is Neither an AI Coach nor a Simple Role-Play Tool

1. Two types of solutions that are often confused with YouMight.

The AI coach

A language model that gives you advice. Available, patient, inexpensive. It can support you on a daily basis and answer your questions. While highly relevant for many use cases, AI coaches have one major limitation, and it is the same limitation as traditional training:

It is exposure to content.

Even with a better interface, the scientific evidence is quite clear: with exposure, people learn, but they don't sustainably change their behavior. Well-formulated advice does not become a behavior. Packaging advice in a chat does not change the physics of transfer.

See our article on the 10% myth.

AI role-play

Role-play tackles this problem at its root: we are finally practicing rather than being exposed to content. It works very well, provided it is as close as possible to the learner's daily reality.

The limitation is that in its usual form, it simulates a conversation:
one person, one conversation, one debrief.

A manager's life is not limited to conversations; it is far more complex. Interactions are numerous, multi-channel, and span days or even weeks.

So while these solutions work in many cases and provide genuine learning solutions, they cannot be sufficient when it comes to management.

2. What makes management difficult is not the conversation.

Let's start with a small Fermi estimate.

A frontline manager, seven people on the team. A typical week:

• 1 team meeting
• ~2 one-on-ones
• 3 to 5 cross-functional meetings
• 40 to 80 messages on internal messaging
• 20 to 40 emails
• around ten unplanned informal interactions

Order of magnitude: 100 to 150 interactions per week, spread across 4 or 5 channels, with 10 to 20 different people.

Mintzberg had already described this in 1973 by observing executives with a stopwatch: short, varied, fragmented work, with a median activity duration of nine minutes.

A real management decision (correcting an employee, announcing a reorganization, arbitrating a conflict between two team members, managing a crisis, etc.) never takes place in a single interaction.

Take the case of handling a sexual harassment situation. This situation unfolds over 2 to 6 weeks, with dozens of interactions: a call with the person who experienced the behavior, informal conversations with various colleagues to understand the situation, formal video interviews with the person responsible, formalization emails, a team meeting where the atmosphere has changed, etc.

A manager's problem is almost never the quality of a conversation. It is the quality of all the actions and interactions carried out around that conversation.

3. So what do we simulate? A company.

YouMight's principle can be summed up in one sentence: we don't generate a person or a situation, we generate an organization.

A YouMight simulation contains:

• An organizational chart (roles, reporting lines, real dependencies)
• Characters, each with their own history, objectives, and good reasons to disagree with you
• A timeline spanning days or even weeks
• Persistent memory
• Multiple communication channels

Enough to genuinely reproduce the life of a manager.

4. Multi-channel, to learn how to navigate.

A manager doesn't just choose what to say. They choose where to say it. And that's often where they make mistakes.

Kruger, Epley, Parker & Ng (2005) measured something interesting: when people are asked to convey a feeling in writing, they estimate that they succeed in ~60% of cases. In reality, they succeed in ~50% of cases. The same messages conveyed orally: ~75% success.

Performance changes radically depending on the channel.

This is in fact the classic finding from Daft & Lengel (1986): the more ambiguous a situation is and the more relationally significant it is, the more it requires a rich channel. Handling an ambiguous issue through a lean channel is a mistake. And it is an extremely common mistake.

At YouMight, channels are therefore part of the exercise: asynchronous messaging, email, calls, group meetings, document sharing. You can decide to resolve the issue with a 3-word message. The platform will let you do it. And three days later, you will understand what that decision produces.

Choosing the wrong channel is a management mistake in its own right. A single-channel simulator cannot make you commit it, and therefore cannot train you to avoid it.

5. Multi-persona, because you don't manage an individual.

A one-on-one conversation is the easy case. What really gets complicated is:

• The team meeting where your ally stays silent.
• The opponent who doesn't confront you directly but influences the entire team.
• The employee who agrees with you face-to-face and then talks behind your back.
• Your manager and your client who have conflicting expectations, with you caught in the middle and no way to satisfy both.

These are distinct skills: maintaining a framework in front of a group, managing a coalition, arbitrating between stakeholders, absorbing public pushback. None of these can be practiced in a two-person dialogue.

YouMight orchestrates multiple characters simultaneously in the same situation, each with their own objectives and their own interpretation of the situation. You are not managing a conversation: you are managing an ecosystem.

6. Self-adaptation, or how to simulate evolving situations.

“That's all very well, but by adding all this complexity, aren't we going to spend hours creating the scenarios? And in real life, things evolve as they go.”

A simulation scenario written by hand requires framing, characters, a branching situation tree, evaluation criteria, and testing. That would not be feasible for the level of complexity we are aiming for.

At YouMight, scenarios are generated. Based on the company's actual context, the manager's role, and what they have actually produced during previous simulations. As they go.

Research on learning is converging: varied practice calibrated to an appropriate level of difficulty durably improves performance compared with repetitive, comfortable practice. This is what Bjork calls “desirable difficulties.”

A fixed catalog gives everyone the same exercise. A generator gives each person the exercise they don't yet know how to do, at the moment they need it.

That's what autonomous evolution means: the platform doesn't deliver content, it creates the desirable difficulty for each person.

Conclusion

Although highly relevant for many use cases, AI coaches and traditional role-play solutions have limitations that make them poorly suited to management.

Managing a team means acting in a complex environment, navigating a system of interactions with memory, channels, divergent interests, and consequences that unfold over time.

A manager who messes up doesn't mess up because they phrased a sentence badly. They mess up because they didn't anticipate what that sentence would trigger two weeks later, in someone who wasn't even in the room.

At YouMight, we are building a platform to de-risk management transitions. Not a chatbot that gives advice, not a role-play with a single character: a simulated organization, with its channels, its characters, its long-term memory, and its scenarios that evolve as you progress.

Sources

• Sharma et al. (2023), Towards Understanding Sycophancy in Language Models — measuring AI assistant sycophancy
• Mintzberg (1973), The Nature of Managerial Work — brevity, variety, and fragmentation of managerial work
• Salas, Bowers & Rhodenizer (1998), International Journal of Aviation Psychology — psychological fidelity matters more than physical fidelity
• Kruger, Epley, Parker & Ng (2005), Journal of Personality and Social Psychology — egocentrism in written communication
• Daft & Lengel (1986), Management Science — channel richness and ambiguity processing
• Bjork & Bjork, Introducing Desirable Difficulties into Practice and Instruction — varied practice, immediate performance vs. durable learning