What AI agents changed about estimation
Estimates used to be expensive guesses that aged badly. When a pre-flight agent sizes every ticket in minutes, the economics flip — and the arguments change.
Estimation has always been delivery’s most reliable argument generator. Teams debate story points like theologians, managers quietly convert them to dates, and everyone performs confidence nobody feels. Deadlines slip anyway — usually for reasons the estimate never contained.
Running an agent fleet changed my relationship with the whole ritual, in three ways I didn’t expect.
Estimates became cheap — so everything gets one
In my pipeline, a pre-flight agent sizes every ticket (XS to XL) with a confidence score before any work starts. It reads the ticket, the codebase, and the recent history, and produces in minutes what a human grooming session produces in an hour of collective suffering.
When estimation costs nothing, you stop rationing it. Every ticket gets sized on arrival, re-sized when requirements change, and the delta becomes information: a ticket that jumped from S to L just told you the requirements moved under it.
Confidence became the real signal
The size was never the interesting number — the confidence is. Low confidence almost always means the ticket is vague, and my pipeline bounces those back for requirements before a line of code exists. That single gate kills the classic failure mode where uncertainty gets resolved silently, mid-implementation, by whoever happens to be typing — the exact spot where audit trails and architecture rules tend to die.
Humans could always have done this. We didn’t, because admitting low confidence in a planning meeting has social costs an agent doesn’t feel.
Appetite stayed human
Here’s what agents did not change: deciding what a problem is worth. Cheap estimates make it tempting to size everything and build whatever’s “only an M”. But cost was never the question — worth is. I still work appetite-first, Shape Up style: decide the budget a problem deserves, then shape the work to fit it. The agent tells me what the ticket costs; it can’t tell me whether the bet is worth making.
If anything, the discipline matters more now. When building gets fast, the scarce resource stops being engineering time and becomes attention — users’ and mine. An accurate estimate of something not worth building is still zero.
The new argument
So estimation arguments didn’t disappear; they moved up a level, to where they belong. We no longer argue about whether it’s a 3 or a 5. We argue about whether it’s worth an afternoon of the fleet’s time — and that’s an argument about the product, not the plumbing.
That’s the quiet pattern in everything agents touch, and it has a name: the mechanical layer collapses, and what remains is judgment.