January 9, 2026 · 5 min read
Estimating AI-Assisted Projects Is Still Hard, Here's Why It Hasn't Gotten Easier
AI coding tools were supposed to make estimates easier by making work faster — instead they made the variance worse, and I've got the missed deadlines to prove it.
I assumed, like most people did, that AI coding assistants would tighten my estimates. Faster typing, faster boilerplate, faster first drafts — surely that compresses the timeline predictably. It didn't. It just moved the uncertainty to a different stage of the project.
Where the time actually went
A feature that used to take four days of writing and one day of review now takes one day of writing and three days of review, because the AI-generated first draft is plausible enough that reviewing it carefully takes real effort, but not so obviously correct that you can skip the careful part. The total time didn't drop as much as the marketing suggested — it just redistributed from a phase I was good at estimating to one I wasn't.
AI coding tools didn't shrink the uncertainty in my estimates, they just moved it from 'how long will this take to write' to 'how long will this take to verify.'
The specific thing that bit me
On a recent integration project, I estimated a data-sync module at two days based on how fast similar modules had gone with AI assistance recently. It took six. The AI-generated code handled the happy path cleanly in twenty minutes — the remaining five and a half days went to edge cases the assistant hadn't considered because I hadn't thought to specify them, which is a prompting failure on my part as much as a tooling limitation.
- I now estimate the 'verify and harden edge cases' phase as its own line item, not folded into 'development.'
- I pad AI-assisted estimates by roughly the same margin I used to pad estimates for unfamiliar codebases, because the unfamiliarity just moved from 'the code' to 'the code I didn't write myself.'
- I stopped quoting clients faster timelines just because the tooling got faster — the review bottleneck didn't move at the same rate.
The tools are genuinely good. But estimating got harder, not easier, because the thing that used to be the reliable bottleneck — typing code — stopped being the bottleneck, and nothing replaced it with the same predictability.