A situation to explore
An unfamiliar codebase becomes easier to extend with AI. A team may ship more quickly while fewer people can explain the changes, diagnose an incident or teach the next colleague how the system works.
The assumption worth revisiting.
This is a workforce issue as much as a technical one. A short-term productivity measure may depend on experienced reviewers whose contribution is invisible. Losing them can change the reliability of the entire arrangement.
If the experienced reviewer left, who could explain and recover what we now produce?
Give the capacity a purpose.
Protect comprehension, debugging and coached ownership when deciding where to redeploy people or absorb a vacancy. Let AI help people learn the system as well as change it.
How this fits the capacity allocation →
Use AI to challenge your first interpretation
Help me explore this question: If the experienced reviewer left, who could explain and recover what we now produce? Identify the assumptions in our current approach. Offer alternative explanations, including one that challenges our preferred answer. Separate hypotheses from evidence and suggest what we should ask the people affected.
Treat the answers as possibilities to test with people and evidence.
Go deeper into the evidence.
Anthropic Research · Research explainer and linked randomised-study preprint
A small experiment recruited 52 mostly junior Python engineers learning an unfamiliar library. The AI group averaged 50% on an immediate comprehension quiz versus 67% without AI: a 17-percentage-point gap. The approximately two-minute completion-time difference was not statistically significant. Explanatory and conceptual uses of AI were associated with stronger understanding than wholesale delegation.
Short task and immediate quiz, not a longitudinal workplace study. Interaction-pattern comparisons are observational. The source's shorthand '17% lower' should be rendered as 17 percentage points using its reported means.
ACM CHI 2025; Microsoft Research · Peer-reviewed survey
319 knowledge workers supplied 936 examples of AI use. Greater confidence in AI was associated with less reported critical-thinking effort; confidence in one's own ability was associated with more. Thinking shifted toward checking information, integrating output and overseeing the task.
Self-reported correlations. This does not prove AI causes declining intelligence or long-term skill loss.
podcast · 33 min 27 sec
From AI experimentation to AI maturity
CIPD
Connects adoption with workforce readiness and returns. Useful when leaders have bought tools but cannot yet identify dependable capacity.
Listen or watch for: Which missing capability is stopping experimentation becoming a repeatable way of working?
About this resource
Practitioner discussion, including a Nationwide representative; it is not an independent evaluation of productivity gains.