I Ran an AI Skill Ranking on My Team (Two Engineers Hadn't Installed Cursor)
I expected to find some 2s and 3s. What I found instead were two people who had never installed the software — nodding along in every meeting while the rest of us talked about workflows they'd never touched.
The first time we ran an internal AI skill ranking, I walked in thinking I already knew the shape of the team. We'd been using Cursor for months. Licenses were paid. The enthusiasts were visible — you could spot them in Slack by who pasted the impressive output. I figured we'd find a cluster of 2s and 3s in the middle, a few 4s, maybe one or two 5s who'd gone fully agentic.
That part was roughly right. What I didn't expect was two engineers sitting at level 1 — not resistant, not skeptical, just never set up. They'd been in every AI discussion, nodded at the right moments, and quietly kept working the way they'd always worked. Without the ranking, they would have stayed invisible until someone noticed their pull requests looked exactly the same as they did a year ago.
I spent the next two weeks with those two people doing nothing glamorous. Installation, orientation, permissions, one real ticket completed with AI assistance end to end. No advanced workflows, no best-practices lecture — just getting from zero to functional. By the end of the second week, they were solid 2s. Without the ranking, those two weeks never happen. You keep running enablement for the people who already showed up.
That's when I stopped treating the ranking as a scoreboard and started treating it as a diagnostic — the thing that tells you who your enablement program is actually for.
Why Teams Skip the Measurement
Most leaders assume adoption because nobody complained. You bought Copilot or Cursor seats, sent the onboarding email, and the loudest voices in the room confirmed that yes, AI is changing everything. Silence from everyone else reads as consent.
It isn't. The ranking breaks that assumption into three populations you can actually plan for: the enthusiasts who would have found the tools regardless, the middle majority who need structure and a nudge, and the invisible baseline — people who haven't started at all. You can't build an enablement program for a team you haven't measured. Most leaders skip the measurement because they're afraid of what they'll find.
That's exactly why you should do it. The average score matters less than the floor. When we ran ours, the team landed at 2.87 out of 5 — uncomfortable for someone who wrote the book on this — but the two level 1s were the finding that changed what we did next.
The Five Levels — What Each One Actually Looks Like
This isn't a performance review. Nobody gets fired for being a 2. The levels describe observable behavior on real work, not enthusiasm or tenure. If people can't recognize themselves in the definitions, the exercise is theater.
Level 1 — Hasn't started
Has not installed or used AI coding tools in a meaningful way. May have a license assigned. May have attended a demo. Has not completed a real task with AI assistance. These people are easy to miss because they're polite in meetings and their ticket throughput hasn't collapsed — they're just working the old way while the conversation moves on without them.
Level 2 — Installed, occasional use
Tools are set up and working. Uses AI occasionally for simple tasks — autocomplete, a basic question, a small refactor. Hasn't adopted a structured workflow. Output is inconsistent because setup varies session to session. This is where most of the middle majority lives if you haven't invested in standards.
Level 3 — Regular use, no system
Uses AI regularly for discrete tasks across the sprint. Can get good results on familiar work. Hasn't integrated AI across the full development cycle — planning, testing, review, deployment. Judgment about when to trust the model is still developing. A lot of "almost right" output that a stronger verification habit would catch.
Level 4 — Consistent workflow
Has a repeatable approach: context loading, mode selection, verification, recovery when the first answer is wrong. Uses AI across most of the development cycle with good judgment about when to delegate and when to think manually. Their pull requests look different from a year ago — not just bigger, but structured differently.
Level 5 — Agentic, teaches others
Operates with a fully agentic workflow. Guides others, contributes to team standards, filters signal from noise on new releases. These people don't need your enablement program — they need a channel to share what works without creating noise for everyone else.
How to Run It Without Creating Anxiety
The ranking fails if people treat it like a judgment. Frame it upfront as a team diagnostic — we're figuring out where to invest coaching time, not who's behind. Self-assessment first, then a short conversation with each person to validate. Discrepancies between self-score and observed behavior are often the most useful data.
Keep the rubric visible during the conversation. Ask for a recent example, not a general feeling. "Walk me through the last story where AI helped" tells you more than "how comfortable are you with Cursor?" Publish the level definitions before anyone scores themselves so level 1 doesn't feel like a confession — it's a starting point with a path.
Run it twice a quarter, not once a year. Skills move faster than annual review cycles, and the point is to catch drift early — especially when a new model or workflow changes what level 4 looks like.
What to Do When You Find Level 1s
Don't send them a tutorial link and hope. The two-week path to level 2 is deliberately boring: get the environment working on their machine, walk through one real ticket together, debrief what worked and what felt weird, repeat on a second ticket with them driving. Advanced workflows can wait. Installation, orientation, and one completed task with AI assistance is the whole goal.
Pair level 1s and 2s with level 5s for weekly working sessions — not a class, a shared ticket. That structure is part of the three-layer enablement system we run now: tiered mentorship, Friday applied sessions on real bugs, and a separate working group for release tracking. The ranking tells you who goes into which lane.
The reframe:
Your AI adoption metrics are hiding your level 1s. Licenses measure access. Rankings measure reality.
What to Do Next
Write the five levels in plain language your team would use in a retro. Run self-assessment this week. Validate with a fifteen-minute conversation per person. Look at the floor before you look at the average.
If you find level 1s, block two weeks for setup and one real task — not a training program, a calendar commitment. The enthusiasts don't need the ranking to tell you who they are. The invisible baseline does.
More field notes on measuring and closing the gap in the blog archive.