Journal

The Pause Nobody Built In

The report goes to the client on a Thursday afternoon. It is good work, or at least it reads like good work. The firm is a consulting practice of about sixty people, and the client is a regional health system that hired it to review how it handles patient billing disputes. A year ago a report like this took two weeks. A senior analyst wrote the first draft, the principal marked it up, the analyst revised, and the principal read it again before it went out. Nobody called that a review process. It was simply how long the work took, and because the report sat on her desk across those two weeks, the principal read it three times.

Now the report takes a day. The analyst works with an AI system that drafts the sections, pulls the supporting sources and formats the footnotes, and the draft comes back clean and confident. The principal reads the summary page, likes what she sees, and sends it. Her calendar is full of other clients, and the firm took several of them on precisely because work that used to take two weeks now takes a day.

Page fourteen recommends that the health system change how it notifies patients about disputed charges, and it rests that recommendation on a state regulation, cited by number. The following Tuesday, before anyone changes anything, the client's in-house attorney does what in-house attorneys do. She pulls the regulation so she can read it for herself. She cannot find it. She searches the state code by number, then by subject, and then she picks up the phone and calls the principal.

That call is how the principal learns there is a problem. She opens the full report for the first time and finds that the citation was invented. So were two others. The client is polite about it, and asks, reasonably, what else in the report it ought to check.

Nobody at that firm stopped caring about the work. Nobody held a meeting and decided the principal would stop reading drafts. The reading went away with the two weeks, and no one noticed it leaving, because no one had ever written it down as a job.

The firm is a composite. I built it from a pattern that has already played out in public at much larger organizations, and smaller versions are playing out right now at companies of every size, most of them with no reporter anywhere near the room.

This is the fourth entry in a series on leading with AI. The Decision You Cannot Hand Off was about the leader alone with a finished answer. Who Signs for It? was about the person that answer landed on, who asked why and found no owner. The Dashboard Is Green was about the system around both of them, and in it I wrote one line I want to build on here: "The handoff between two people put a second set of eyes on the problem." This entry is about what happens to that second set of eyes when the slow work that supplied it goes away.

Look at what the two weeks were buying the firm. The first thing was a second reader. The person who makes the work tends to read what they meant to write, while the person who did not make it reads what is actually on the page. The second thing was time. A citation that looks fine on Monday can look strange on Thursday, when you come back to it fresh and something about the number snags. The third thing was a last point where a mistake was still cheap. If the principal had caught the invented regulation on her third read, the cost would have been an hour of the analyst's time.

None of those three jobs appeared on an org chart or in a process document. They lived inside the delay. I wrote about this in RemAIning Human, and the sentence still holds: "But speed removes friction, and friction has always served a purpose." The firm bought speed and got it. What it gave up in the trade was a check it never knew it was paying for, so it never thought to keep it.

Three organizations have learned this in public, and in each one AI produced the work.

The first is Deloitte Australia. In December 2024, Australia's Department of Employment and Workplace Relations hired the firm to review the IT system the department uses to automate penalties in the country's welfare system. The contract was worth about A$440,000, and the 237-page report was published in July 2025. Chris Rudge, a welfare academic at the University of Sydney, read it closely and found papers attributed to his own colleagues that did not exist, along with a quote from a federal court judgment that the court never wrote. He counted up to 20 errors in the first version, and he told reporters that his colleagues' work had been cited but not read.

Deloitte published a revised report that disclosed the use of a generative AI system, Azure OpenAI, and agreed to repay the final installment of the contract. To be fair to Deloitte, the firm did not answer when reporters asked whether AI caused the errors, said the matter had been resolved directly with the client, and says the corrections do not change the report's findings. The department said Deloitte's own review confirmed that some footnotes and references were wrong. What the record does not show is who read the report before it left, and I will not guess at Deloitte's internal process. A firm hired to give the government an independent review delivered 237 pages, and the reader who caught the problems was an academic on the outside.

The second is a summer reading list. On May 18, 2025, the Chicago Sun-Times ran a special section called "Heat Index," and inside it was a list of fifteen summer books. Ten of them did not exist. A freelance writer produced the section for King Features, a syndication company owned by Hearst, and he told reporters he used AI for the list and did not fully vet what it gave him. The Sun-Times said the section was licensed content that its newsroom neither created nor approved. The chief executive of its parent company wrote that the circulation department, which worked with King Features to bring the section in, expected it to match the standards of Chicago Public Media. Two months earlier, 20 percent of the paper's staff had accepted buyouts. That is context, and I am not offering it as the cause. The writer had a gate, the syndicator had a gate, and the newspaper had a gate, and each one assumed someone else was standing at it.

The third is Zillow, and here AI made the decision itself. Zillow Offers bought homes directly from sellers, pricing its offers with the company's home-value model, the Zestimate, and a seller who accepted had a deal. A case study in the Journal of Information Systems Education describes how it started: after the model produced a value for a home, Zillow's pricing experts vetted the bid and adjusted it based on their understanding of market conditions. Then Zillow began using the Zestimate itself as the cash offer on qualifying homes. The same case study reports that the company then regularly prevented its pricing experts from modifying the model's values. In the third quarter of 2021, Zillow took a write-down of about $304 million, because it had paid more for homes than it now expected to sell them for. On November 2, 2021, it announced it would wind the business down and cut about a quarter of its workforce. Zillow's chief executive explained it this way: "We've determined the unpredictability in forecasting home prices far exceeds what we anticipated." You can weigh that explanation for yourself. An unpredictable market is exactly where a person who knows the street earns a place in the loop.

That is where leadership failed in all three cases. In each one, the work crossed from the people who made it to the people who would act on it with no one else reading it first. The AI produced what it produced. Language models invent citations and pricing models misjudge markets, and anyone who has used either for long knows it. The failure belongs to the organization, which let the output leave at a speed or a volume no person was checking. Somebody decided to adopt the tool. Somebody decided to take on more clients, or more homes, or more syndicated pages. Nobody decided who would read the work before it reached the people who would act on it, because that job had always come free with the slow pace, and a job that comes free has no owner. When the slowness went, there was no one whose role it was to defend the check.

I wrote in Accountability Is a Design Problem that goodwill alone does not sustain accountability, and that healthy systems design it in. The same is true here. What these organizations lacked was a place for care to happen, built on purpose. The fix is friction in one place, chosen deliberately.

I call that place the gate. The gate is the last point before work leaves the people who made it and lands on someone who will act on it. A report goes to a client, a page goes to readers, an offer goes to a seller. Inside the gate, a mistake costs a redo. Past it, a mistake costs a refund, a firing or a write-down. That point is where a leader should spend the friction.

Putting a reader at the gate takes five decisions, and every one of them is a leadership decision about how the organization works. None of them is a setting on a tool. The leader decides which work counts as leaving the building. Most internal email can move at full speed, while a client deliverable, a public statement, a price, or a policy that lands on customers or employees needs a second reader. Then the leader names the reader, one person who did not make the work, so the job belongs to someone. Then the leader puts protected time on that reader's calendar, because reading squeezed in between other things quietly stops happening. Then the leader hands the reader two plain questions to bring to everything that crosses: What in this could be false? Who acts on this, and what happens to them if it is wrong? Finally, the leader gives the reader the authority to hold the work, and backs them in public the first time they use it.

That last decision matters most, and the research explains why. A sign-off box on a form is a gate with nobody standing at it. In 2010, Raja Parasuraman and Dietrich Manzey reviewed the research on how people work alongside automated systems. They found that people lean on automated output whether they are novices or experts, that training and instructions do not prevent it, and that complacency appears when attention is split across several tasks. In 2021, Zana Buçinca, Maja Malaya and Krzysztof Gajos tested designs that forced people to stop and think before they accepted the AI's answer. Compared with simply showing people the AI's explanations, those designs significantly reduced overreliance. They did not eliminate it, and people rated the most effective designs the least favorably.

Put those findings side by side. A reviewer with a full queue becomes a rubber stamp, however careful that reviewer is, and the structure that works is the one people like least. That is why time and authority belong to the gate. Without them, you have built a signature line and called it a check.

The first objection is that this is bureaucracy. Bureaucracy is friction everywhere, applied out of habit until people learn to route around it. The gate is friction in one place, on the work that costs the most when it is wrong, and everything else keeps the speed you paid for.

The second is that your people are careful. I believe you. Careful people work at big firms and good newspapers too, and the research says careful people with full queues stop checking. That makes this a question of capacity, and capacity is something a leader can give.

The third is that the reader will become a bottleneck. A reader with a queue and no authority does become one, and soon signs things unread just to keep up. A reader with protected time, two questions and the power to hold the work becomes the reason the work can be trusted.

The fourth is that AI keeps getting better. It does. Zillow considered its model a major strength, and the market still moved in ways the company said it could not forecast. A better tool increases how much work crosses the gate. Someone still has to be standing there.

So here is the work for this week, whatever your role. If you lead a company, the gate is yours to set up. If you lead a team, it is yours to set up inside your part of the company. If you are the one who hits send, it is yours to ask for.

Start by listing what your team now sends out faster than it used to. For each item on that list, name who reads it before it leaves, other than the person who made it. Some of those lines will come up empty, and that is the finding. Pick the empty line that would cost the most if the work were wrong. Name the reader, block the time on their calendar, and hand them the two questions. Tell the team that the reader can hold the work, and back them out loud the first time they do, even if the work turns out to be fine, because that first hold is when everyone learns whether the authority is real. If you do not hold that authority, take this step to the person who does. Last, say plainly what this costs in speed and why you are paying it, so the slower turnaround reads to everyone as a choice.

The first entry in this series asked what the output does not know about this person. The second asked whose name is on it. The third asked what we stopped hearing. This one asks something just as plain: what do you let out of the gate this week that nobody but its maker has checked?

Go back to the firm. The report still takes a day. The analyst still works with the AI system, and the draft still comes back clean. This time, before it goes to the health system, it goes to a senior consultant who did not write it and who has an hour on her calendar for exactly this. She reaches page fourteen, stops at the regulation, and goes to look it up. She cannot find it. She walks the report back down the hall, the analyst fixes three citations, and the report goes out a day later than it could have. The client's attorney reads it the following Tuesday, finds every regulation where the report says it will be, and moves on to the next thing on her desk.

If you are working out where the gate sits in your own organization, I would like to hear about it. Let's have a conversation. It's free.

Footnotes That Exist

This post had help from AI, and the AI got things wrong. It first brought me cases that were not AI failures at all. It put a researcher in the wrong field. It placed three authors at a university where only one of them works. It repeated a line about a company's model that no source supported. Even in the writing of this post, the AI I built to help with research had several significant misses that wouldn't have been caught without me remaining in the loop. That is this post's argument, applied to the post itself.

Deloitte Australia

Chicago Sun-Times and King Features

Zillow Offers

The research