The AI coding adoption percentage on your CTO's slide isn't a productivity number

The myth

If 64 percent of our engineers are using a coding agent and saying they’re at least 25 percent faster, the AI investment is working and the WAU chart is the proof point.

That’s the myth I want to retire this week. It’s clean, it’s quantitative, it survives a board meeting, and it is almost entirely a license-utilization story dressed up as a productivity story.

Why it sounds right

I get why it lands. A CFO can read a 64 percent number. A board can read a 25 percent velocity bar. An adoption chart is the rare AI artifact that looks like the kind of operating metric companies have been pricing for decades.

It’s also the kind of number every harness vendor would love your CTO to bring to the next QBR. Cursor publishes context-usage breakdowns. Claude Code surfaces session hours. GitHub Copilot exposes 28-day active-user dashboards down to active-versus-passive code review. The instrumentation is genuinely better than it was twelve months ago. The dashboards are real. The percentages are real.

The trouble is that the percentage is measuring whether the harness was opened, not whether the company shipped something it can defend in 30 days.

What the evidence says

Look at the Jellyfish 2026 State of Engineering Management report, published May 7 alongside coverage in DevOps.com. Same survey, 636 software development professionals fielded in March. Headline number:

That’s the slide your CTO is bringing. Now the next slide, from the same report, same respondents, same survey instrument:

Two 64 percents on the same page. The first says “we are faster.” The second says “we cannot prove it.” Both are true at the same time, which is the most honest thing I have read about coding agents this quarter.

Three more numbers from the same survey are worth the price of admission:

• Only 53 percent say AI has actually improved the quality of the code being committed. The other 47 percent are either neutral or negative on quality, which is a different way of saying half the room is shipping faster into a verification problem.
• 42 percent flag the increasing cost of AI tools as a top concern. That tracks with what I am seeing on the spend side this week. Anthropic doubled Claude Code’s five-hour rate limits on May 6 and Knightli covered the rollout on May 9. The ceiling is rising because the heavy users are pushing through it, not because everyone with a seat is consuming evenly.
• 36 percent report senior engineer reluctance to adopt. Which means the engineers with the most judgment about when to trust AI output are the underweight slice of the WAU number a CTO is putting on the slide.

There’s a second piece of evidence worth the detour. SD Times’ May 8 weekly roundup walked through Coder Agents going to beta, Opsera-Cursor, Snyk-Claude, Prismatic Skills for Claude Code. Read the announcements together and a pattern emerges: the enterprise harness market this week is not selling a better adoption number. It is selling the control plane that catches what the adoption number is hiding. The reason inboxes are full of governance-layer pitches right now is that the people on the inside know the WAU chart on its own is not a board-defensible answer.

The reframe

Adoption is a license-utilization metric. Productivity is a delivery metric. Conflating them is the same mistake we made with lines of code in 2014 and dashboard-driven OKRs in 2019. The harness made the mistake easier to repeat because the dashboards are slick this time.

The board-defensible number is not how many engineers opened Claude Code or Cursor last week. The board-defensible number is some version of “verified output per engineer per month, net of the verification work the rest of the team had to do downstream.” That number lives in the merge log, the incident rate over a 30-day window on agent-touched code, and the senior-engineer review queue. None of those numbers are on the WAU slide.

Adoption is a license-utilization metric. Productivity is a delivery metric. Conflating them is the same mistake we made with lines of code, on glossier dashboards.

The cleaner mental model: a harness adoption percentage tells me how much optionality the engineering org has installed. A productivity number tells me how much of that optionality the org converted into shipped, verified work. The first is a ceiling. The second is a floor. Most CTO slides this quarter are showing the ceiling and calling it the floor.

So what

Three things to ask your CTO this week, before the next board prep.

First, ask for the same slide with the verification-confidence number sitting next to the velocity number. If 64 percent of the engineers say they are faster and 64 percent of the engineers say they cannot confidently measure it, those two bars belong on the same chart at the same height. Right now they almost never do.

Second, ask for the senior-engineer adoption split. The Jellyfish survey said 36 percent of orgs are seeing senior reluctance, which is the cohort whose judgment most predicts whether agent output ships safely. If a high-WAU number is being driven by junior and mid-level engineers while the staff-plus tier opts out, the headline number is reporting a different question than the one a board is trying to answer.

Third, ask for the cost-per-merged-PR trend over the last 90 days, with token spend included. Anthropic’s rate-limit ceiling moving up this week, the Opsera and Snyk control-plane partnerships, the cost concern showing up at 42 percent in the Jellyfish data, all three signals point to the same thing: variable AI spend is becoming a budget line that needs its own scorecard, not a footnote on the productivity slide.

None of this is a reason to slow down. Most of the engineering orgs I am talking to right now have real, durable gains from the harness rollout. The 25 percent velocity number is probably directionally correct for the teams that are paying attention. The problem is not the gains. The problem is the slide. If the board is going to make a hiring-plan or budget call in the next quarter on the strength of one chart, the chart should be the one your team can still defend in October.

The companies I find genuinely impressive on this right now are not the ones with the highest adoption percentages. They are the ones whose engineering leaders are quietly putting the second 64 percent number on the same slide as the first.