We put AI on every desk and left the room that decides empty

For the last two years we've pointed almost every AI dollar at the doing level of the company, the desks, the agents and the private decks, on the assumption that faster output downstairs would eventually unblock the room upstairs. It didn't, because the room that decides still runs the way it ran before the chat window showed up, and every slide we can now produce in minutes gets fed into the same six-to-ten-week meeting cycle that was already failing.

McKinsey's State of AI 2026 makes the split hard to ignore. Eighty percent of respondents say AI improved their own productivity, while only 37 percent say their company saw any EBIT impact at all and just 6 percent work at companies earning 5 percent or more of their EBIT from it. Those aren't steps in a funnel, they're answers to two different questions, and the distance between them is the part of the company we left empty.

The pile nobody puts on a dashboard

Take one ordinary operating question. A company of about 800 people has two support offices, West wait times are the problem, and the COO asks whether to combine them. What follows is a first meeting where two executives are traveling, weeks of teams building slides so their bosses look prepared, a vote at the third meeting and Legal and HR hearing about it afterward, which means everyone meets again. Six to ten weeks later the fourteen people whose jobs change are still guessing, and the wait times haven't moved.

None of this is new. Bain found in 2006 that only about 15 percent of the 365 companies it studied were highly effective at making and executing decisions, and in 2014 Mankins, Brahm and Caimi traced one weekly executive meeting that set 300,000 person-hours a year in motion, of which only 7,000 were the meeting itself. The chat window didn't create that pile, it just made every piece of it cheaper to produce. Microsoft's 2025 data shows the heaviest users now get interrupted every two minutes during core hours, and Gallup puts manager engagement worldwide at 22 percent in 2025, down from 31 percent in 2022, so the people staging these meetings are nearly as checked out as everyone they lead.

Why more AI conversation makes it worse

The instinct is to throw more agents at the problem, but a finance agent isn't a prompt that says "you're the CFO." Sharma and colleagues showed at ICLR 2024 that language models drift toward whatever the user already believes, so when you ask the same model the East-West question ten different ways, the answers collapse toward the easy option with the least conflict. Taubert and colleagues found at ACL 2025 that longer discussion before a vote reduced performance, and in a March 2026 preprint, Prakash reports that a structured argument beat a free-for-all on hard calls, especially where facts only one participant holds decide the answer.

Structure is the variable, not more voices. It also has limits worth stating plainly: the same preprint found no gain on routine work, the strongest evidence comes from studies of models rather than executive teams, and a file is only as honest as the sources each seat brings, so this belongs on the handful of calls that shape the company rather than on every purchase order.

What the decision layer actually needs

A question someone can sign. "Support strategy" is a topic, and topics produce decks, while "Do we combine East and West support?" is a sentence someone can put their name under, which tells you who has to sit and what a finished answer looks like.

Seats, not costumes. Each function carries its own data, the limits of its job, a record of what it argued last time and a goal it can only break on the record, so finance attacks cash while operations attacks what happens when one office becomes the only office.

Two passes and a neutral middle. The first pass surfaces the facts only one seat holds and the second attacks the options that survived, while someone in the middle, human or agent, admits only requested evidence and can't carry a version of the story.

One file instead of a deck. The document holds the recommended answer, the vote by seat, the losing side in its own words, the claims that got stripped and the reopen facts, and it goes out before the meeting.

One meeting, then work. The only question in the room is whether to sign, any "why not" has to point at a page, and after the signature the tools you already run, agents included, work from the file instead of from a hallway sentence.

That last move is the one AI projects skip. RAND found in 2024 that 84 percent of the industry interviewees it spoke with named leadership decisions as the root cause of AI project failure, and MIT's NANDA project found that only 5 percent of organizations got a custom enterprise AI tool into production, because tools asked to invent a decision stall while tools that inherit one ship.

What it's worth

McKinsey's 2021 transformation survey found that about 45 percent of the value lost in a transformation is gone before implementation even starts, while the company is still arguing about the target, which is exactly the stretch this compresses.

The labor math is my own and deliberately a floor. At 250 hours the old way and 25 through a file, priced at a $95 loaded hourly rate built from Bureau of Labor Statistics data, one decision drops from about $23,750 in labor to $2,375. Over a year, a company of 800 that signs 24 decisions through the file instead of finishing 8 the old way gets back 1,400 hours and roughly $133,000, and with the board signing four decisions a quarter it lands around 40 signed calls a year, against 12 in the old way's good year.

The bigger change is the kind of miss you're left with. Today the miss is internal, because nobody owned the answer. Decide first and whatever miss remains belongs to the market, where a call was made, signed and simply wrong, which is a failure a company can actually learn from.

We had an engine built for exactly this work, holding every fact and keeping every killed option visible, and we used it to write email.