Two different roles for AI in the room
There are two ways AI can enter a high-stakes decision. One is as a stand-in for the room itself: model the stakeholders, predict their reactions, rehearse the decision against the model before you make it for real. The other is as a working partner to the decision-maker's own reasoning, something you consult, argue with, and consciously integrate, without letting it quietly take over the read you're capable of doing yourself.
Wethos AI is betting on the first. My work, and the framework I teach executives, System 1-2-3, trains the second.
These aren't mutually exclusive in principle. But in practice, when a leader outsources the "what will they think" question to a model, something gets skipped: the harder, slower work of learning to read a room, hold a disagreement, and notice your own blind spots in real time, with an AI tool folded into that process rather than substituted for it. That skill doesn't transfer from a rehearsal against a simulated stakeholder. It's built the way any perceptual skill is built, through repeated, reflective contact with the real thing, with feedback that sharpens the instrument rather than replacing it.
What System 1-2-3 actually trains
System 1-2-3 is a decision-literacy framework, not a decision-support tool. It starts from the well-established distinction in behavioral science between fast, intuitive judgment (System 1) and slow, deliberate reasoning (System 2). System 3 is the layer most frameworks skip, and it isn't a third mode of human cognition alone. It's the deliberate skill of working at the intersection of your own biological decision-making, System 1 and System 2 together, and artificial decision-making: the outputs, predictions, and models an AI system puts in front of you. System 3 is the trained habit of bringing the two together on purpose, instead of letting one quietly stand in for the other without you noticing it happened.
Most executives are skilled System 1 operators. That's how they got the job. The failure mode isn't a lack of intuition, and increasingly it isn't a lack of access to AI either, it's not knowing when to distrust either one. A reorg decision that feels obviously right in the moment, or a prediction an AI tool hands you with high confidence, are exactly the kind of inputs where System 1-2-3 diverges: the leaders who get burned aren't the ones who lacked information or lacked a model. They're the ones who never developed a disciplined way to hold AI input in the room without letting it quietly replace their own read of it.
That habit is trainable. It's also portable in a way a stakeholder model isn't: it travels with the person into the next decision, the next room, the next AI tool that didn't exist when they learned the skill.
Where modeling the room actually helps, and where it doesn't
I want to be fair to what Wethos AI and tools like it are good at. If you need to pressure-test a specific announcement against a specific, well-documented set of stakeholders (a board you've worked with for years, whose reactions you have real data on) a model can catch a blind spot before it becomes a live-fire mistake. That's genuinely useful, and I wouldn't tell a client not to use it as one input into their own System 3 process.
But it has a ceiling, and the ceiling is the model's inputs. A model of your CFO built from past meetings, memos, and stated preferences will miss what your CFO hasn't said yet, hasn't fully worked out herself, or would only reveal to a person she trusts in an unscripted moment. High-stakes decisions are disproportionately made in exactly those unscripted moments, with people whose positions are still forming. No model, however well permissioned, has access to a mind that hasn't finished making itself up.
And there's a second cost that's easy to miss: every time a leader reaches for a simulated rehearsal instead of doing the harder work of preparing themselves to read the room live, the underlying skill atrophies a little further. Decision-support tools that make the leader's own judgment less necessary tend, over time, to make it less sharp. That's not a criticism unique to Wethos AI. It's the general pattern with any tool that substitutes for a skill instead of training it, and it's worth naming plainly because the industry mostly doesn't.
The augmentation test
Here's a simple test I use with clients when they're evaluating a new decision-support tool, whether it's a modeling platform, a dashboard, or an AI copilot: does using this tool, six months from now, leave you better at the underlying judgment, or just more dependent on the tool for it? That's the question System 3 is built to keep asking, every time an AI output enters the room.
Tools that pass this test show you their reasoning, invite you to disagree with it, and get out of the way once you've internalized the pattern. Tools that fail it get more indispensable the longer you use them, because they were never designed to make themselves unnecessary.
System 1-2-3 is explicitly designed to pass its own test. The goal of every session is a leader who needs the framework a little less than they did before, because the underlying instrument, their own calibrated judgment at the intersection of intuition, deliberation, and whatever AI is in the room with them, has gotten sharper. That's a different business model than a platform built on stakeholder data and subscription modeling, and I think it's worth executives understanding the difference before they decide which kind of capability they're actually trying to build.
A concrete case
Take a reorg announcement, a common trigger for exactly this kind of tool. A leader using a stakeholder-modeling platform would feed in what's known about each direct report, generate predicted reactions, and rehearse the announcement against the model's best guess at each person's response. If the model is well built, this can genuinely soften a rollout: fewer surprised reactions, a script that anticipates the loudest objections.
Now run the same reorg through a System 1-2-3 lens. Before the meeting, the leader isn't asking "what will each person say," they're asking a prior question: is my confidence in this reorg coming from System 1, a pattern-matched sense that it's obviously right, or has it actually survived System 2 scrutiny, the kind where I've tried to argue against my own plan and it held up? That question alone catches a category of mistake no stakeholder model can catch, because the mistake isn't in anyone's predicted reaction. It's in the leader's own reasoning, before a single person in the room has said a word.
System 3 shows up a beat later, when the leader runs the same plan past an AI tool, a scenario model, a synthesized stakeholder summary, and has to ask a further question: is this output telling me something my own read actually missed, or am I about to let it stand in for a read I haven't done yet? That's the discipline the framework trains: not avoiding AI input, but knowing how to hold it at arm's length long enough to interrogate it, before deciding whether it earns a seat at the decision.
In the meeting itself, the two approaches diverge again. A leader working from a rehearsal script is watching for confirmation that the room is behaving the way the model predicted. A leader trained in System 1-2-3 is watching the room itself, noticing in real time when someone's body language doesn't match their words, when a question that sounds procedural is actually the objection, and switching consciously into slower, more deliberate listening the moment the stakes in the room shift. That's a live skill. It has no script to fall back on when the room does something the training data never anticipated, which, in a genuinely high-stakes reorg, it usually does.
The room you're actually in
The next high-stakes decision you face won't come with a permissioned model of everyone in it. It will come the way they always do: incompletely, with people whose minds are still changing, in a room where the read has to happen live, AI tools available to consult but never to defer to by default. You can walk in with a rehearsal script built from last quarter's data, or you can walk in with an instrument you've spent real time sharpening, one built to work with AI rather than be quietly replaced by it.
I'd rather help executives build the second thing. It's slower to build and it doesn't scale the way software does. But it's the only one of the two that's still working when the model runs out of data, which in the rooms that matter most, is usually right when you need it.