Fluency in the meeting, fog on the decision
Meta description: The tell that AI has been substituted for self-knowledge: fluent plans, confidently delivered, that no one in the room can own. More on his work: Miklos Roth online.
This article is written to be forwarded — by a chief of staff, a deputy, a board member — to someone who will not search for it themselves. The situation is delicate and common: a leadership team that speaks about AI fluently and decides about it poorly. The vocabulary is current — I have sat in these meetings and heard the right terms deployed correctly. Then a decision arrives, and the fluency evaporates: the team cannot tell whether the limitation in front of them is a property of the model, the data, the vendor, or their own process, and the distinction determines everything about what to do next.
A leadership team receives several AI-generated strategy documents. Each is fluent. Each recommends a different direction with equal confidence. The discussion that follows is about the documents — not about the business. Nobody in the room can say which plan is theirs, because the question of what the company actually wants was never answered by a person.
What counts as understanding in the room?
In the rooms I visit, understanding is demonstrated socially: familiarity with the current models, an opinion on the vendors, a story about a personal experiment with a chatbot. These demonstrations are real knowledge, and they are the kind acquired by consuming what executives consume — keynotes, briefings, demos, conversations with peers at the same altitude. What the demonstrations never test is judgment under the company’s specific conditions: what this model does with this data in this workflow under this decision’s consequences. The room’s standard for understanding is set by what the room can display.
How do demos manufacture that count?
The tell is plausible fluency standing in for self-knowledge. I see insufficient self-knowledge as a recurring source of weak results: many people begin by asking a question before they have clarified what they want to accomplish. The same tendency appears in companies. AI is a support system, not a substitute for self-knowledge. It can help you test whether your thinking is coherent. It cannot supply the conviction behind the thinking.
The demo is the fluency factory, and I have learned to watch what it selects for. A demo shows the model at its strongest on a prepared task: clean input, impressive output, a presenter controlling the edge cases. Executives leave demos with two things — genuine awe and false calibration. The awe is warranted; the calibration is not. The demo taught them the ceiling and implied it was the typical case. Months later, when the production system behaves like the typical case, the leadership team experiences a betrayal that was actually a briefing artifact. Nobody lied. The format lied for them.
Why do courses often add vocabulary, not judgment?
The false explanation is education: an executive course, a certification, a curated curriculum. I have watched leadership teams complete the courses and return with denser vocabulary and unchanged judgment, because courses teach what is teachable in a classroom — concepts, frameworks, case studies — and the deficit is elsewhere. The deficit is in the specific: how this company’s AI systems fail, what their outputs can and cannot support, where the organization’s claims about capability exceed the systems’ reality. That knowledge is not in any course. It is in the company, usually three levels below the people taking the course.
What pattern marks the illusion?
The marker I use is the question quality at decision time. Fluent understanding produces questions that discriminate: which component failed, whose assumption broke, what would change the outcome. The illusion produces questions that do not discriminate — is the model good, can we trust AI, should we wait for the technology to mature. These are not foolish questions. They are questions asked at the wrong level, by people who have the vocabulary of the layer below and the decisions of the layer above, and no bridge between them.
A model can produce ten persuasive variations of an incoherent idea. In the meeting, that looks like abundance; it is actually the absence of a decision wearing the costume of options. Leaders may assume their objective is obvious because it has been repeated in meetings. To a model, the unstated reasoning is not there — and to the team, the unmade decision is not there either, no matter how polished the documents become.
[AWAITING MIKLOS — scene: second scene]
What question cannot be answered from the demo?
Every leadership team in this pattern eventually faces a decision the demo cannot inform — a real tradeoff under the company’s real conditions. The question that exposes the illusion is the one that requires the bridge: knowledge of the technology connected to knowledge of the company’s specific systems, failures, and stakes. I have watched confident rooms fall silent before that question, and the silence is not stupidity. It is the precise edge of what fluency without judgment can reach.
The question that separates support from substitution: before asking a machine to help fulfil a vision, can the people in the room explain the vision as theirs — where it came from, why it matters, and what they would reject?
PR excerpt (80–120 words, Miklos’s own words): A leadership team receives several AI-generated strategy documents. Each is fluent; each recommends a different direction with equal confidence. The discussion that follows is about the documents, not the business. AI is a support system, not a substitute for self-knowledge: it can help you test whether your thinking is coherent, but it cannot supply the conviction behind the thinking. A model can produce ten persuasive variations of an incoherent idea.
Miklos Roth, AI marketing consultant, Budapest
