A measuring device, with its faults marked
This one measures a person. It places you on six levels using facts you can check — not statements you agree with — and then names the single thing standing between you and the next one.
It is assembled from three separate findings that had never been put together: a stage ladder whose value sits in the middle, a two-axis structure where the axes compound, and one worker's narrated climb with observable markers along it.
This is a defensible composition, not a validated instrument. Its ladder shape is borrowed from an enterprise model built on a 2022 survey. Its two-axis structure is transposed from an organisational construct measured across 3,467 respondents. Its level markers come from a single narrated case — self-reported, unaudited, and selected for being worth broadcasting — plus one population-scale measurement of learning effects.
That is enough to locate yourself and pick a next move. It is not enough to rank people, gate a promotion, or put a number in a performance review. Treat a result as a conversation starter with yourself.
v2 changes, after scoring three real cases. Five defects in v1 were found by running it against people rather than reasoning about it: one gate conflated a practice with an artifact and sat a level too high; one gate asked you to rate a capability instead of report a behaviour, and could collapse the whole ladder on its own; one gate's wording covered scheduled work but not event-triggered work; the two axes were measured with 4 questions against 9; and two gates were near-duplicates. All five are fixed. The full record, including the cases and what changed because of which, is in the calibration report.
v3 changes — the fold. Two editions of this instrument were built independently on the same night: this gate engine (local session) and a six-dimension profile with a cited action layer (remote session; 17-source evidence base, filed as syntheses/ai-worker-maturity-levels in the wiki). v3 folds them. The fifteen calibrated gates are unchanged word for word; from the remote edition come the visual identity, an L0 Bystander floor (one new screening gate, G0, excluded from the axis scores), source citations on every gate action and level plan, answer persistence in this browser, and the METR self-report caveat. Two stale v2 strings ("Thirteen gates", "of 14 answered") are corrected. The calibration record and the scored derivation case document the v2 engine, which is carried here intact.
Why two axes and not one
The strongest argument against a single score comes from a 3,467-respondent study that scored organisations on two learning capabilities at once. Either one alone helped a little. Having both compounded — up to 2.2× on some measures. The majority, 59%, were low on both. The axes below are that finding transposed to a person.
Axis F
Can you get good work out of the model, and — the part that actually matters — can you tell when it hasn't? This is measurable and learnable: after controlling for task type, model choice and complexity, longer-tenured users show systematically higher task-success rates.
The terminal skill on this axis is not directing. It is evaluating — knowing what good looks like precisely enough to check the work.
Axis S
Have you restructured your own work so the AI has context, reach, and somewhere to put its output? This is where the compounding happens, and it is why someone can write excellent prompts all day into a stateless chat window and still capture a fraction of what's available.
Most of the observable markers live on this axis, which is why the assessment asks more questions about it. Facts about your setup are checkable; feelings about your fluency are not.
The assessment
Answer from what is true this week, not what you intend to set up. Every question asks about something observable — the surest way to get a useless result is to answer aspirationally. Each level is a gate, not a score: you are at the highest level whose gates you have all passed.
The AI Worker Maturity Scale.
Result
PDF opens your browser's print dialog — choose "Save as PDF" as the destination.
All 16 gates and the diagnostic, as answered
| Gate | What it tests | Your answer | Status |
|---|
Provenance
Nothing here is invented. Each part of the model is traceable to a page in the wiki, and the full audit trail — including the 65 pages the query considered and rejected, with a reason for each — sits alongside this file.
MIT CISR's four-stage enterprise model (28 / 34 / 31 / 7%), and its finding that the greatest financial impact comes between stages 2 and 3 — not at the top. sources/2026-04-28-mit-sloan-ai-maturity
Ransbotham et al., MIT SMR × BCG, n=3,467: two learning capabilities scored independently, compounding when combined, 59% low on both. sources/2026-05-07-ransbotham-augmented-learners
The corpus's only narrated end-to-end worker trajectory: share of work in one harness, who writes the system of record, agent-detected context gaps, friction logging, distribution. sources/2026-08-31-blum-how-i-ai-claude-cowork-pm-system
Anthropic Economic Index 5: high-tenure users show higher task-success rates after controls, and model selection tracks task value. sources/2026-05-07-anthropic-economic-index-5-learning-curves
"Delegate tasks, not judgment." Google's Developer Intelligence team on the five patterns of top AI-native engineers, and verification as the new bottleneck. sources/2026-04-21-forsgren-macvean-build-core-skills-thrive-ai-era-developer
The productivity J-curve at individual scale: return runs negative first, trust falls, you double-check everything. A snapshot cannot tell a climb from a failure. concepts/micro-productivity-trap
Judgment and evaluation, not tool knowledge — evaluation named the terminal skill for people managing fleets of agents. concepts/durable-skills
A company of Level-4 individuals is not a Level-4 company. Distribution turned out to be a UX problem: three days by terminal failed, fifteen guided minutes worked. concepts/enterprise-ai-adoption
All four maturity instruments in the wiki measure firms; success rates span 1–13%. The worker-level half of that thread's question is now answered by the wiki synthesis; external validation is still wanted. threads/ai-maturity-measurement-comparison
The per-gate citations (Mollick, Karpathy, Argenti, Osmani, Beutler, Carson, Kiron & Schrage, METR) come from the remote edition's 17-source evidence base, fused as the wiki's worker-maturity synthesis. syntheses/ai-worker-maturity-levels