Why more AI output is making most teams slower

Start with the clearest version of it. An internal update that used to be one slide is now five. The request did not change, the audience did not change, and the decision the deck exists to support did not change. What changed is that producing five slides costs the author roughly what producing one used to cost, so five is what comes back from the model and five is what gets sent.

The author saves a few minutes. Everyone downstream pays for four slides that carry no decision. Multiply that across every deck, every specification and every summary of a summary, and the company is reading more than it was a year ago without knowing more than it did.

The mechanism is not complicated. Generation got cheaper. Comprehension did not move at all. A model produces a specification faster than anyone can read it, and the review of that document still runs at human reading speed, in a meeting, with people who have to hold enough of it in their heads to disagree with it.

So the constraint moved. It used to sit with the person producing the artifact. It now sits with everyone who has to read, check and act on what was produced, and almost nothing about how we work has moved with it.

The volume concerns me less than what it replaced. The dangerous pattern is the good first draft. Something plausible arrives, it reads well, it is better structured than the author would have managed on their own, and nobody performs the step where they decide whether they agree with it.

In finance that gets expensive fast. A finance user will not act on a recommendation they cannot interrogate, and they are right to refuse. Under SOX or the EU AI Act, a decision whose reasoning nobody can reconstruct is a decision the company cannot defend. That is the standard we hold our own agents to: every action Aimie takes is traceable back to the data behind it, because a collections decision has to survive an audit. I would apply the same test to a person who forwarded a model's output without reading it.

This next figure is an impression from walking the floor rather than a measured number: roughly one person in ten is now much faster than they were, and the rest are producing more.

Everyone has the same tools. What the faster group does with them is specific enough to describe. They use the models to read. They arrive at the meeting with the one slide, having had the model go through the five. They know the rough error rate on their own kind of task, because they checked, so they know which outputs to trust and which to re-derive. And they point the tools at whatever is actually constraining them. If the bottleneck is a review queue, generating more documents to put into it makes the quarter worse.

The tempting measures are the ones the tools hand you: documents produced, tickets closed, commits, tokens spent. Every one of them rises when nothing is happening.

The measures worth having are the ones the business already had, and in Order-to-Cash they cash out as money. DSO. Cash application match rate. Manual touches per invoice. How long a dispute stays open. We are our own client zero, so those are the same numbers our customers hold us to, which makes them difficult to argue with internally.

I do not yet have a clean before-and-after on any of them that I would attribute to AI adoption. That is my gap, and I would rather say so than publish a productivity figure I could not defend in front of the people who produced it.

The measurement I want next is the one nobody publishes: how much of what we produce is read, by whom, and what they did differently afterwards. Until I have it, one slide and five slides look identical on every dashboard I own.