The Estimate Is the First Thing AI Broke

I gave an estimate recently that my gut was completely sure of, and it was wrong by a wide margin, because the gut producing it was calibrated to a world that has quietly been retired. The strange part was not being wrong, everyone is wrong about estimates sometimes. The strange part was how certain the wrong number felt on the way out of my mouth. It carried exactly the same weight, the same settled confidence, as a good estimate. Nothing about the feel of a reliable estimate and an unreliable one has changed, even though the ground underneath both of them has shifted completely.

That is the specific trouble. The instinct still fires with full confidence. It just fires at a target that moved, and it cannot tell that the target moved, because an outdated instinct feels identical from the inside to a current one.

Estimation is pattern-matching to "last time something like this took about this long." It works as long as the relationship between effort and time stays roughly fixed. AI broke that relationship, and it broke it unevenly, which is the part that actually matters. Some work collapsed to minutes, the generating, the boilerplate, the first draft of almost anything. Other work did not move at all, and some of it got slower: the understanding, the review, the integration, the part where a human being has to actually absorb and take responsibility for what the machine produced.

So the old average is now meaningless, because it was an average across two things that used to move together and no longer do. Your gut hands you one confident number, and that single number is silently blending something that got ten times cheaper with something that got no cheaper at all, and sometimes got worse. The blend comes out feeling precise. It is not precise. It is two very different numbers wearing a trenchcoat.

I stopped estimating the thing as a single number. I estimate the parts separately, because only some of the parts got cheaper and lumping them together is exactly what produces the confident wrong answer. The question I ask is no longer "how long will this take." It is "how much of this is generating, which is now fast, and how much is understanding and checking and integrating, which is not." The second number is the one that actually sets the timeline now, and it is precisely the number the old instinct discounts, because it used to be bundled invisibly inside the first and never had to be named.

The move that has helped most is small and slightly uncomfortable. I say the estimate, and then I say out loud what would make it wrong. The estimate as a hypothesis with a tripwire attached, not a promise I have to defend. "Two days, and here is the specific thing that, if it turns out to be true, makes it five." It keeps the number honest, and more importantly it gives everyone in the room permission to revise it the moment the tripwire trips, instead of quietly defending a figure that stopped being true the instant the work actually started and the ground shifted again.

That is a less comfortable way to estimate than the old confident single number. It might also be the only honest one available right now, given how fast the thing underneath the estimate keeps moving.

So when you give a number now, are you estimating the work, or are you estimating the version of the work that stopped being true a few months ago?