On mature codebases at large enterprises, a project that took six weeks a year ago might take about five weeks today. That can look like the entire return on millions of dollars per month spent on AI coding tools.
AI agents make coding fast enough that it stops being the obvious bottleneck. The bottleneck shifts to validated decisions about what customers want, sanity-checking AI’s work, slow rollouts, customer feedback, and deciding what to do next.
Six weeks of wall clock time becomes five weeks at best, not six days. People read this as a bear case for AI productivity. My argument is that they're measuring the wrong thing — project duration is not the only, or even the primary, place where AI productivity shows up.
What has changed for us is that we have many more projects running in parallel, we're raising the bar across the board, and we're working on codebases that we never touched before. I think these are the right places to look for AI productivity in engineering teams in 2026.
1. Run more projects at once
We run two or three projects in parallel per person, and people like it more because they're less stuck waiting on a human review. The constraint becomes how fast a PM or designer can carve out independent chunks of work. It still takes six weeks to ship one project, but instead of three engineers doing one project for six weeks, one engineer can oversee three projects in that time. In this comparison, engineering labor efficiency improves ninefold even though project duration barely moves.
2. Raise the bar
In my experience, an average PR costs roughly four hours of wall clock time whether it's sloppy or polished. So we ship the polished one. Real monitoring, end-to-end tests, delightful animations — work that used to rot on the backlog doesn't need to go there when an agent can do it right away.
3. Widen every engineer's range
Teams without iOS engineers used to watch mobile drift behind web. Now an engineer can often open the iOS codebase, make a change, run the simulator, and ship it. The same is increasingly true for new service setups, migrations, infrastructure work, and design polish. Specialists still matter, especially for high-risk work, but product engineers can cross boundaries much more readily. The baseline for "full-stack" keeps expanding.
So measure more than cycle time. Track how many validated projects a team can sustain, what quality bar it can maintain, and how safely people can work across codebases. AI's productivity gain is as much an expansion of capacity as a compression of time.