Golden Data

Issue #21 · September 30, 2026

The AI Uplift Is Already in the Roadmap. The Measured Gain Is Eight Percent.

Delivery dates are being set on a speedup almost nobody has measured.

Paul BrownWrites Golden Data from Ogden, Utah5 min read

The 2027 roadmap goes up for sign-off for most leaders about now. The delivery dates in it already assume the build team is faster than it was last year. Nobody wrote that assumption down. It lives in the gap between what got committed and what the team shipped.

DX, which sells engineering-productivity measurement to about 500 companies, put a number on that gap on April 28. Across more than 400 engineering organizations tracked from November 2024 to February 2026, AI tool usage rose 65% and median pull-request throughput rose just under 8%. The data comes from DX’s own customers, so take the direction, not the decimal. In the same report, DX says the leaders it works with are being held to 3x or 10x, and that coding is about 14% of a developer’s week. A faster keyboard moves a small slice of the work. Eight percent is a real gain. It is not the gain in the plan.

The delivery date is the one input in the plan nobody has to expense. A tool needs a budget line. A hire needs a req. A date needs a meeting. So the AI uplift got priced in the only place it was free to go, and it lands on the analysts and engineers who own the dates. The person who built the capacity plan is the person who explains it in Q2.

What’s actually shipping this week

None of these is new this week. The planning calendar is what renews them: the roadmap gets committed now, and each signal below sits under one of its dates.

May 13, Harness’s State of Engineering Excellence 2026. 700 developers and engineering professionals at enterprises. 81% of engineering leaders say the time AI saves on coding is now spent reviewing AI’s output. Nearly a third of a developer’s day goes to that review, and none of it shows up in output metrics. The saving is real. It was spent before it reached the roadmap.

August, LeadDev’s AI Impact Report 2026. 600 engineering leaders. 70% have adopted AI developer tools widely or completely. 26% report major productivity gains. Separately, only 31% measure those gains at all. The claim and the measurement are not the same group.

September 2025, DORA’s State of AI-assisted Software Development. A year old and still the independent baseline in this list. 90% adoption. AI is now associated with higher delivery throughput and, still, with higher delivery instability. A date set on the throughput inherits the instability.

Now the counter-read, and it is a real one. DX’s own benchmark refresh on July 21 shows systems-measured PR throughput up nearly 20% year over year, with a visible jump after a February model release. If that holds, the 8% is already stale and the gains are compounding. Then the uplift is real. It is real for the teams on the new tools with the new habits, and a flat raise in commitments charges every other team for it. One number cannot tell those two cases apart. A line item can.

What I’d build in the analytics org this week

An AI-uplift line in the 2027 capacity plan, backed by a backtest, so the speedup has to earn its place before a date is promised.

Planning moment: the 2027 roadmap commit and the capacity plan under it.

Data shape: Jira or Linear issue history and GitHub pull requests by team and sprint, joined in BigQuery to the AI vendor’s admin usage export by person and week, into fct_team_sprint (team, sprint, committed_items, shipped_items, cycle_time_days, pr_count, review_wait_hours, rework_rate, ai_active_share, tool_rollout_date). Shipped means in front of a user. Merged is not shipped.

System shape: a Looker explore over dbt-governed definitions, with two cohorts: teams where most members were active on the tool for a full quarter, and teams where they were not. Compare shipped items and cycle time before and after the rollout date, cohort against cohort. Put review wait and rework beside them, because Harness and DORA both say that is where the saving leaks. The measured gap is the uplift line. The assumption becomes a cell with a source.

Eval shape: the gap in shipped items, not in PRs. If PR count climbs and shipped items stay flat, the uplift is in the keyboard, not the roadmap. Say so in the plan.

Latency budget: before the roadmap is committed; a one-time run, then quarterly.

Cost shape: one analyst-week. The usage export and the ticket history are already paid for.

Instrumentation: keep the rollout dates and the usage export as tables, not screenshots. Self-reported time saved is not a source. METR’s July 2025 trial found experienced developers 19% slower with AI while they believed they were 20% faster. Next year’s plan reads the table, not the survey.

Two weeks in: success is a capacity plan with AI uplift as its own row, a number, and a source, set beside the headcount assumption it used to hide inside. Credible failure is ticket history too inconsistent to define “shipped.” Then the row says “unmeasured” in plain text, and the dates carry last year’s velocity.

Pull it out as its own line. If it cannot earn the row, it is not a delivery date. It is a hope with a deadline.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown