Blog
Shaping the future of how engineering teams work, ship, and scale
Recent Posts
Cost per merged PR: measuring the cost half of AI coding ROI
Divide a quarter's AI coding spend by the human-authored PRs merged in it, per team. That's what you paid per shipped change, not your return. AI makes PRs cheap, so pair it with the all-in cost per PR, the rework share, the kind of work and the significance of what shipped.
How to measure ROI after deploying an AI coding cost tool
Measure an AI coding cost tool by the decisions it causes, not by total spend. A baseline, what counts as savings, a 30/60/90-day plan and a worked example.
How to measure the ROI of AI coding tools
Measure AI coding ROI by team: count every cost, measure the return as significance-weighted shipped work, and report the break-even lift the tools have to clear. A worked example for a 60-engineer company.
How to report engineering ROI and capacity allocation to the board
Build the board's engineering update around what shipped and why it mattered. Frame it with capacity allocation against target, give cost as context with AI tooling on its own line, and close with risks and asks.
What software measures the ROI of AI coding assistants?
No single product measures AI coding ROI on its own. Here's what the vendor dashboards, engineering intelligence platforms, session cost tools and work-outcome tools each measure, whether they can see the return side, and what to ask before you buy.
How to set a 2027 AI coding budget your CFO will sign
Build a 2027 AI coding budget from what last quarter's tokens bought: spend by team and kind of work, the share lost to rework, and next year's roadmap. A worked example with arithmetic you can check.
Why the average AI coding spend per engineer misleads
AI coding spend is heavily skewed, so cost per engineer and per-seat math mislead. Read the distribution instead: median vs. mean, the top sessions, and what they were for.
Cap, alert or coach: how to govern coding-agent spend
Govern coding-agent spend with three levers: cap, alert or coach. A cap bounds the bill, an alert catches runaways, coaching finds what spend bought.
How coding agents bill teams, and how to compare them
Coding agents now bill teams by usage, in credits, pools and allowances, so seat price says little. Compare tools on your own work: one reference price, the same kind of work, and what each session produced.
Tokenmaxxing Counts the Input. Measure What the Tokens Bought
Tokenmaxxing is running up AI token use to look productive. Tokens are an input. Measure what they bought: cost per PR, rework spend, cost by kind of work.
Adoption Rate Tells You Nothing About Whether AI Is Working
Engineering leaders watch AI adoption metrics climb: 40% to 60% to 80% of engineers using Copilot daily, throughput up, cycle time down. The numbers are real and the question they answer is wrong. Adoption measures diffusion. It does not measure whether the rollout is actually working.
Code Review Used to Be How Craft Propagated
For a generation, engineering orgs kept what their best engineers knew by passing that knowledge through code review. AI agents moved the quality-determining decisions upstream, out of any artifact the org can see. The instrument of knowledge transfer is still in place. The knowledge is no longer passing through it.
Cognitive Debt Is a Session Problem, Not a Documentation Problem
Engineering teams are shipping AI-generated code their engineers cannot explain. The industry has named this correctly (cognitive debt) and prescribed the wrong fix. Documentation requirements and review checkpoints ask engineers to develop comprehension the session didn't create.
The Output Was Never the Work
For a decade, an industry agreed that pull requests, commits, and cycle time were the right units for measuring engineering. That premise held until AI broke the proxy. When a developer can generate ten pull requests in an afternoon, the artifact count is identical to what a highly productive human produced in a week, and the dashboard cannot tell them apart.
Speed Was Never the Gap
Harvard-backed research confirms it: AI tools make developers faster. That speed is not converting to better software at the same rate. The instinct is to diagnose this as a measurement problem and reach for better analytics. That diagnosis is wrong.
The Case Against DORA Metrics
DORA metrics promise to measure engineering effectiveness, but they're optimizing for the wrong things. Here's what you should measure instead.
Introducing Maestro AI
We built Maestro because engineering leaders deserve better than vanity metrics. Here's our vision for impact-driven engineering intelligence.
The Engineering Intelligence Revolution
Engineering leadership is evolving from gut feel and vanity metrics to AI-powered intelligence that actually understands your team.
Why Activity Metrics Fail
Measuring lines of code and commit counts is like judging a novel by word count. Here's why activity metrics fail—and what to measure instead.