• A diagnosis, not a score. Don’t rank teams by level. Compare a team with its own earlier state.
• “Not yet” means not yet shown. It is not forbidden: a level 3 team’s gains can last, and a level 4 team can question its strategy. Work on whatever limits you now.
• The gain decides. AI that takes the initiative is the usual route. A team that starts its AI routines by hand is at the same level if the gain is the same.
• Moving up takes more than one change. A first step shows whether you are on the right track. The level needs its condition to hold across the unit’s work, with evidence.
• Mixed and unknown are valid answers. Place the most important work separately. Don’t average it into one number.

A real check follows one piece of recurring work: the useful result and who uses it, the old way, what AI did, the time period, all the effort (drafting, checking, fixing, waiting, switching), costs pushed onto others, and the failures too.

People and AI working together are part of how the organization does, coordinates and improves its work. The levels measure what that changes in results, not how much AI is used.

Skill in using AI, sometimes called taste, matters at every level: knowing what good work looks like, giving AI a useful task, checking it against sources, and knowing when to stop it. People retain their own judgment and the ability to handle consequential choices and failures.

Attention counts as a cost. Count session switching, rebuilding context, interruptions and supervision. Agents can reduce these burdens or add to them. Check the effect wherever the routine is used.

• The number of tools, agents or prompts, or how much people use AI.
• How autonomous the AI is.
• Another AI agreeing with the first. Check against sources, requirements and tests.
• Documents, dashboards and meetings that do not change results.

Mara and her team are fictional, a composite of people who lead new-product work.

These studies shaped what the levels look for.

• METR (2025): experienced developers felt about 20% faster with AI and were 19% slower.
• Dell’Acqua et al. (2023): big gains inside AI’s capability, worse answers just outside it.
• Lee et al. (2025): more confidence in AI went with less critical thinking.
• Bastani et al. (2025): assisted output is not the same as learning.
• Brynjolfsson, Li & Raymond (2025), Generative AI at Work: real gains in one firm, with wide variation.
• Vaccaro, Almaatouq & Malone (2024): people plus AI is not automatically better.
• Mark, Gudith & Klocke (2008): interrupted people work faster, with more stress.
• Brynjolfsson, Rock & Syverson (2021), The Productivity J-Curve: investment is not realized productivity.
• Wageman, Gardner & Mortensen (2012): team effectiveness includes future capability.
• OECD (2001), Measuring Productivity: output relative to all resources used.