We've done itbackwards

It starts small, with the CEO, the COO and the CFO agreeing that the whole executive team should hear it. A meeting goes on the calendar a week out, two executives turn out to be traveling and staff starts building slides, and because almost nobody has the full picture, that first week isn't a decision at all, it's a hunt for numbers and a deck.

Then the rest of the company gets pulled in. Teams meet, those meetings spawn more meetings, and hundreds of hours go into making the next executive meeting look prepared, until the team presents, nobody signs and someone suggests giving it two more weeks for comments. A third meeting goes on the calendar with more slides and more people catching up.

At the third meeting they vote around the table, while the two who missed the first meetings are still reading last week's deck. Legal never sat in those rooms and neither did HR, so they hear about the call afterward, and because they have things to say, everyone meets again.

By now it's been six to ten weeks, and the teams are guessing their jobs from a decision they only half heard, while the fourteen people whose roles change still don't know what happens to them. The wait times haven't moved either, and a customer who was already waiting too long didn't stay for the third meeting.

Someone asked a tool on Sunday and someone upstairs asked a different one on Monday, so twelve decks arrive looking ready, each one built to survive a boss rather than to finish the decision.

The tools turned out to be good at preparing the pack and nearly absent from the conversation in the room, and that split is the whole bet in one sentence. We assumed more output downstairs would unblock the room upstairs, but it fed the same bottleneck, so people still decide slowly and then software goes fast on a fuzzy sentence.

What got governed was the keyboard, after the fact, while the decision itself never got governed at all.

Nobody owns the answer, too many people can stop it, and the meeting ends first.

That isn't a 2024 finding. Crozier saw layers of management keep their power by keeping a little fog in the system, and Stasser and Titus watched groups talk through the shared slides while the one fact that would have changed the call sat unspoken in someone's folder, which is how a group picks the popular option and calls it judgment.

The structure rewards exactly the wrong behavior. It pays people to protect turf, to reopen a call because they missed the meeting, to treat a veto as if it were a control and to call harmony alignment, and adding size doesn't fix any of it, because another steering committee is just a better place for the empty chair to hide.

A 2023 study in the Journal of Strategic Information Systems described two loops running at once: a top-down loop that can line a company up, and a bottom-up loop that protects local interest. The bottom-up waste is real, but it's usually a rational response to nobody owning the answer upstairs.

The meeting isn't the hour on the calendar. It's email before six, an interruption every two minutes, a deck rebuilt while people walk in and the inbox open again at ten, which is what "give it two weeks" actually buys.

Managers used to be the engaged ones.

Managers used to sit above the people they lead as an engagement premium, and that premium is gone, because Gallup now finds them nearly as checked out as everyone else, except in the companies that still treat a manager's week as work instead of deck production.

It's the manager, not the license, that decides whether AI actually gets used. Gallup's February 2026 survey of 23,717 US employees found manager support was one of the top two drivers of frequent AI use, yet fewer than one in three employees strongly agree their manager actively supports it.

Now put a price on one decision.

The COO's question doesn't add a new tax, it rides the one the company already pays, because six to ten weeks of slides is that same interruption stack pointed at a single question. Take the hours from the East-West clock, price them at what the Bureau of Labor Statistics says those seats cost to employ and keep the arithmetic honest, which makes the result a floor that real mid-market pay packages sit well above.

The leak isn't a mystery.

Value dies while the company is still arguing about the target, dies again while people try to execute a fuzzy sentence, and dies a third time after someone stamps the initiative done. The people who'll actually do the work spend those months manufacturing private versions of the story so a manager looks briefed, and none of that activity is the decision.

Reverse the sequence and nobody builds slides to impress the room, because the room already has the document. The executives stop hosting a tour and start doing the one job only they can do, which is signing the file or sending it back.

Different jobs, not one voice in ten hats.

A finance seat isn't a prompt that says "you're the CFO." Ask the same kind of model a question ten different ways and the answers collapse toward agreement, because agreement is what these models are trained to produce. If you run the East-West question ten ways on one pile of documents, the group lands on the easy option by the second pass, because that's the option with the least conflict and the fewest reasons to reopen. Sitting right next to that fake agreement is the other failure, which is invented evidence that makes the agreement look sourced.

A real seat holds things a costume doesn't: its own data and the limits of the job, plus a history of what it argued last time and what happened next, along with a goal it can only break on the record. Advocacy is the whole point, because polite generalists produce mush.

On East-West, finance has to attack cash while operations attacks what happens when one office becomes the only office, and the West lead files the wait times that only West holds. If they all sound like the same reasonable person by lunch, you don't have a debate. You have a press release.

Two passes, each with a job. More talk makes it worse.

In a 2026 preprint, Prakash reports that a structured argument beat a free-for-all on hard calls, with the biggest gains where facts only one person holds decide the answer, while the same system didn't help on routine work, so don't aim it at office chairs. The warnings point the same way: Taubert et al. (ACL 2025) saw performance drop when groups discussed longer before voting, Wynn and Song (2025) watched accuracy fall the longer models kept talking, and Sharma et al. (ICLR 2024) showed models drifting toward whatever the user already believes.

So the passes aren't more conversation. The first pass surfaces the facts only one seat holds while the second attacks the options that survived, and the checkpoint between them is the only door new facts can come through, with someone who takes no side standing at it.

Someone in the middle who doesn't take a side.

Passes without a checkpoint turn into chat, and chat is exactly what those papers say fails. The middle looks hard at the file, asking whether the group surfaced facts only one person holds or just recited the shared slides, and whether anyone flipped without a reason or let the easy option slip through unmarked. Then it writes down the options, who fought whom, the vote by seat, the citations that survived, the claims that got stripped and the data that's still missing.

That person can't carry their own version of the story, because the moment they do you've rebuilt the old staffer and quietly created a second decision-maker. The seats argue and the middle compiles, while the person who owns the outcome signs.

The meeting gets smaller. The record stays.

The meeting stops being "what do we think" and becomes "do I sign, and if not, why not," and why not has to point at the document: a missing source, a flip without evidence, a limit the signer chooses to break or a reopen fact that's already true.

That's where the hours come back. Mankins' pile is mostly people manufacturing a version of the story before anyone decides, and one shared document turns all of that into a read, a signature and a handoff. After the signature, any meeting that isn't doing the work or hitting a reopen fact is a leak.

When a call goes off the rails, companies hire someone expensive to reconstruct what was known and who's safe to blame, and the file makes most of that archaeology unnecessary because operations said no, cited the risk and watched the COO sign anyway, all on the page. Two rules keep that record honest. Nobody gets punished for being on the losing side since that would train the next group to be peacemakers, and reopen facts are the only legal way back into a decision.

Stop quoting the 70 percent.

Beer and Nohria wrote in 2000 that about 70 percent of change efforts fail, and they had no study behind the number. When Hughes walked the citations back in 2011 he found no valid base for a universal failure rate, so this paper sticks to what was actually measured.

What was measured points in the same direction anyway. The people building AI systems keep naming leadership failures rather than technical ones: the wrong problem, the wrong measure of success, weeks demanded for work that takes months and priorities switched halfway through. Tools that inherit a clear way of working make it to production, while tools that don't inherit one stall, which is why the model is so rarely the thing that broke.

Today the miss is internal, because nobody owned the answer, the wait times stayed where they were and a customer left while the company was still in slides. Decide first and the miss that's left belongs to the market, where you combined the offices and service still failed or the customer simply said no. That company lost in public, while the other one never decided at all.

What changes with the size of the company.

Twenty people. In a company this size the founder is the room, the board and the only memory of why the last call went the way it did, so when that person's on a plane the whole company is working in fog. The file becomes the company's memory with jobs attached to it. Use it on the five or ten questions a year that actually are the company, like pricing, a senior hire, a customer you fire or a raise, and new people can read the file instead of rebuilding the company's history from lore.

Two hundred to two thousand. This is the East-West firm, big enough for people to hide inside functions but not big enough to pretend that combining two offices is somebody else's regional problem. Start with the executive team, because that's where the pile is heaviest and where the hours come back first. Legal and HR stop showing up as the fourth meeting, and managers get back a week that's actual work instead of slide production, which is the engagement lever Gallup has already measured.

Five thousand and up. Don't talk about "the whole enterprise," because a company this size doesn't have one cascade, it has a standing pile, and it won't start deciding in days just because a new tool landed on the CEO's calendar. The move is to flip the questions a unit, a region or a product line genuinely owns and stop sending them upstairs as tours, so the enterprise benefit shows up as a portfolio of written questions rather than one heroic transformation.