In April I argued that most AI in a company is pointed at the people, not at the business. Every person gets faster. The wait between them stays the same. I called it being stuck in 1992, and I left it as an argument.
This is what happened when I timed my own loops.
The claim, in one picture
A company runs on a loop: build, ship, measure, decide, build again. AI shrinks the boxes. It has done almost nothing to the arrows. If you want one test, pick something your team shipped last month and count the days until you knew whether it worked, then the days until you changed something because of it. That sum is your company's real speed. I have never seen a token budget move it.
I run a small fleet of agents that write, publish, and measure content for me, with me at the approval gates. It is the closest thing I have to a company where every loop is visible. So I went and measured the arrows. Two of them.
A thirty-day arrow I did not know I had
By May, an agent was drafting the LifeHack newsletter and it went out twice a week. Fast box. The measure step was still me, opening the dashboard when I remembered to, which turned out to be rarely.
I know exactly how rarely, because the store keeps dates. Six newsletters went out between June 4 and June 23. All six were measured on the same day: July 4. The oldest had waited a month for anyone to look at a number about it. The drafting agent kept producing, on schedule, with no idea whether the last one worked.
On July 9 I closed the loop. A measure agent scores each send on human clicks once it is a few days old, and the manager that kicks off the next draft reads those numbers first. Nothing about the writing changed. Since then every send has been measured within four to six days, and the result is in front of the next run before I have opened anything. A month became under a week, and I am still the only human in the loop, at the same two gates as before.
That is the whole thesis in one system. The making was already automated. The waiting lived on the measure side, and nobody had looked there because nothing was broken.
A sixty-second arrow that behaves like a month
The second one is smaller, and it shows the mechanism. The dashboard for AgentUse, the open-source agent runtime I build, has a little dot that shows which agents are running. It was lying. Runs started from the command line showed up as live 12 to 18 seconds late. Worse, finished runs kept showing as live for 45 to 60 seconds after they had ended.
Cosmetic, you might say. Except that dot is a measure step. If it says live when the job is already done, whoever is watching waits, and the next decision waits with them. Sixty seconds on my dashboard. Days when the dashboard is a monthly report. A wrong signal is not a neutral signal. It adds a queue you cannot see.
The root cause was a cache that only refreshed on events the dashboard itself owned, so anything that happened in another process was invisible to it. I got both edges down to about two seconds, and along the way corrected two claims I had made about the code once I measured instead of trusting my reading of it.
Automate the measuring before the making
Neither fix made anything write faster. Both shortened an arrow, and both arrows were on the same side of the loop.
That is the rule I have landed on. An AI-shaped company is not one where everyone uses AI. It is one where every loop has something automated on the measure-and-feed-back side, not just the do side. A fast maker with a slow measurer produces more untested work, not more learning. I wrote once that open loops forget and closed loops compound. The measure side is where the loop actually closes.
What to do on Monday
Pick one loop. Not your whole stack, one loop. Write down how many days pass today between shipping the thing and a human seeing a number about it. Then put your first real automation on that side, the side that tells you what happened, not the side that produces more stuff.
Expect the first thing you learn to be that your measurement was wrong. Mine was.
Faster people were the easy part. The company is next.