Your Leadership Team Already Disagrees. The Meeting Is Hiding It.
Ask a chief executive what their executive committee decided last month and you will usually get a clear answer. Ask four members of that committee the same question separately and the answers diverge more often than anyone is comfortable with. Not because people are dishonest, but because the meeting that produced the decision was not built to surface disagreement. It was built to get through an agenda.
This is the most expensive unexamined process in most large companies. Everything downstream of it, including every AI programme, inherits its quality.
The Pattern in One Paragraph
A pre-read goes out late and is skimmed by roughly half the room. The agenda is ordered by function, so items arrive in the order the org chart implies rather than the order of contention. The first person to speak anchors the discussion. Consent gets confused with agreement, because silence is indistinguishable from either. Items with genuine disagreement are the ones that run out of time, so they are deferred, and the deferral is recorded as progress. The decision that emerges has an owner in the minutes and no owner in reality.
Why Agendas Are Ordered Wrongly
Almost every executive agenda is assembled the same way: each function submits its items, and the items are sequenced by seniority, by rotation, or by whoever asked first. None of those orderings has anything to do with where the decision risk actually sits.
The consequence is predictable. Items where the room already agrees consume live discussion time, because they are easy to talk about and produce a pleasant feeling of progress. Items where the room genuinely splits arrive at minute fifty of a sixty-minute meeting, when the only available move is to defer. Over a quarter, this produces an executive team that has thoroughly discussed everything uncontroversial and repeatedly postponed everything that matters.
An agenda ordered by contention rather than by function inverts this. Items where the room already agrees collapse to a line of information. Items where it splits get the time. The change sounds administrative and is not: it alters what the leadership team is actually for.
The Disagreement Nobody Says Out Loud
The harder problem is that the split is frequently invisible. Executives are experienced people who read the room, and reading the room is exactly the behaviour that destroys the information the room needs.
The mechanisms are well documented and predate AI by decades. Irving Janis named groupthink in 1972 — the tendency of cohesive decision-making groups to suppress dissent in favour of unanimity. Two decades earlier, Solomon Asch's conformity experiments had already shown that people will contradict clear physical evidence to match a group majority. Cass Sunstein and Reid Hastie later catalogued how information cascades — where each speaker updates on what was said before rather than on what they independently know — and group polarisation make committees systematically worse than their members (Wiser, Harvard Business Review Press, 2015). Kahneman, Sibony and Sunstein's work on noise added a further point that most boards have not absorbed: noise here means unwanted variability between decision-makers assessing identical information, and that variance is consistently far larger than the decision-makers themselves believe.
None of this is a character flaw. It is what happens when positions are formed in public and in sequence. The fix is structural, not motivational: collect positions privately, before the discussion, and let the aggregate be visible before anyone speaks.
| Conventional exec meeting | Instrumented exec meeting | |
| Positions formed | In the room, in sequence | Privately, in parallel, beforehand |
| Agenda ordered by | Function or seniority | Degree of disagreement |
| Consensus items | Discussed at length | Collapsed to information only |
| Contested items | Deferred when time runs out | Surfaced first, with the split visible |
| Silence means | Ambiguous | Recorded position |
| Decision exits with | A minute entry | A named owner and a date |
Where AI Belongs Here, and Where It Does Not
There is an obvious temptation to point a language model at the meeting transcript and ask it what was decided. This is the wrong end of the process. By the time there is a transcript, the anchoring has happened, the quiet dissent has gone unrecorded, and the model is summarising a conversation that already lost the information you wanted.
The useful application sits earlier and is narrower than most vendors suggest. Ahead of the meeting, a brief goes to each participant and each participant gives a short private read on each item. What follows should be split carefully:
How much the room splits on an item, which items are contested, who is isolated on a position: these are counts. They should be computed deterministically and be reproducible. A model that estimates them will occasionally be confidently wrong, and one confidently wrong claim about a named executive's position ends the tool's credibility permanently.
Clustering free-text comments into themes, drafting the brief, turning a decision into readable follow-up: these are genuine model tasks where an imperfect result is recoverable.
The system should not tell the executive team what to decide. It should tell them where they are not aligned, and give them the time back to resolve it. Tools that recommend decisions get switched off within two cycles.
That division of labour is not a technical detail. It is the reason the output can be shown to a chief financial officer without a caveat, and it is the same principle we described in what AI-native actually changes in your operating model.
The Switch-Off Test, Applied to a Meeting
The test for whether a process is genuinely redesigned rather than merely accelerated is to switch the system off for a cycle and see what happens. Applied here, it is unusually clarifying.
If the tool disappears and the executive team goes back to a function-ordered agenda, an unread pre-read and a decision log nobody owns, then nothing was redesigned. A summarisation layer was added to an unchanged ritual. If the tool disappears and the team still collects positions before the meeting, still puts the contested item first, and still refuses to close an item without an owner and a date, then the operating model changed and the software was the delivery mechanism.
The second outcome is the only one worth buying, and it is not primarily a software outcome.
Why This Is an Embedded Problem
Executive decision-making is close to the worst possible candidate for a procured product, for reasons that have nothing to do with the technology.
The ritual is specific to the company. The politics are specific to the company. The set of things that can be said out loud is specific to the company, and it is not written down anywhere. No demonstration environment contains any of this, and no requirements document survives contact with it, because the requirements that matter are the ones nobody will put in writing.
What works instead is the pattern we set out in how to buy AI delivery that actually ships: someone sits inside the real meeting cycle, week after week, and adapts the instrument to the room rather than asking the room to adapt to the instrument. The tell that this is happening is that the product changes shape in the first month. The tell that it is not is a configuration call and a training session.
It is also why this work stalls under a conventional programme structure. There is no function that owns "how we decide", so there is nobody to sponsor fixing it, which is the failure mode we described in why innovation programmes stall without operators.
What to Do Before Buying Anything
Three diagnostics, none of which require a vendor:
Separately, in writing, within 48 hours of the meeting. The spread in the answers is your baseline, and it is usually the number that starts the conversation properly.
Go back through a quarter of minutes and mark every item deferred more than once. Those are your contested items, and the pattern of them tells you what the meeting is systematically avoiding.
Not an accountable function. A named person and a calendar date. In most executive logs the honest figure is under half, and closing that gap requires no technology at all.
If those three diagnostics come back healthy, you have a well-run executive process and you do not need software for it. If they come back the way they usually do, the problem is real, and it is worth being precise that what you are fixing is a process rather than shopping for a meeting tool.
Want the Disagreement Surfaced Before the Decision?
We build instruments like this inside the real meeting cycle rather than alongside it, scoped to one decision process and finished when your team runs it without us.
See how we work →Sources & References
- Irving L. Janis, Victims of Groupthink (Houghton Mifflin, 1972; 2nd ed. Groupthink, 1982) — the original account of how cohesive decision-making groups suppress dissent. Source for the claim that the mechanism predates AI by decades.
- Solomon E. Asch, conformity experiments (1951 onward) — source for the claim that people contradict clear evidence to match a group majority.
- Cass R. Sunstein and Reid Hastie, Wiser: Getting Beyond Groupthink to Make Groups Smarter (Harvard Business Review Press, 2015) — source for the information-cascade and group-polarisation claims, and for committees underperforming their members.
- Daniel Kahneman, Olivier Sibony and Cass R. Sunstein, Noise: A Flaw in Human Judgment (Little, Brown Spark, 2021) — source for the claim that variance between decision-makers assessing identical information is larger than they believe.
- Daniel Kahneman, Dan Lovallo and Olivier Sibony, "The Big Idea: Before You Make That Big Decision…", Harvard Business Review 89(6), June 2011, pp. 50–60 — the 12-question pre-decision checklist; source for the principle that structured checks belong ahead of the decision rather than after it.
- Practitioner observation, SUPALABS engagements 2024–2026. The three diagnostics, the design constraints (two minutes, no typing) and the adoption patterns described are drawn from our own client work. This is unpublished, proprietary and not independently verifiable — treat it as informed opinion rather than evidence, and weigh it accordingly against the sourced claims above.
Questions fréquentes
Innovation9 min2026-07-29EN
Partager cet article
