L5 · 05 · Leading AI Adoption in Teams
The last skill isn't technical: it's getting a team to actually use AI well, measuring whether it helped, and setting policy people follow. Most AI rollouts fail not on the tech but on adoption — no training, no measurement, unrealistic promises, or a ban that pushes people to shadow AI. This page is the playbook.
By the end of this page you can
- Run an AI adoption playbook: baseline, pilot, train, scale, measure.
- Measure productivity honestly — without falling for the vanity metrics.
- Design training that sticks (pairing, workflows, champions).
- Write a sane AI usage policy people will actually follow.
- Handle the human concerns: skill atrophy, job anxiety, uneven adoption.
Why rollouts fail
| Failure mode | What happens |
|---|---|
| Buy licences, declare victory | Tools installed, nobody's workflow changes, usage decays to zero |
| Hype the numbers | "30% faster!" with no baseline — credibility dies at the first bad output |
| Ban it | People use personal devices/accounts with no controls — worse than governed use |
| Mandate it | People perform compliance, resentment builds, gains don't |
| Ignore the humans | juniors stop learning, seniors fear displacement, quiet resistance |
The adoption playbook
- Baseline first. Before any AI, measure how long key tasks take and their quality. Without it, "AI saved us time" is unprovable.
- Pilot small. One willing team, real work, a few weeks. Find the workflows where AI genuinely helps.
- Train on workflows, not features. "Here's how to review a PR with AI" beats "here's what the chat box does".
- Scale with champions. Early adopters who share real wins (and failures) teach better than any deck.
- Measure against the baseline. Honestly — see below.
Measuring productivity honestly
The hard truth about "X% faster"
Vendor studies claim huge gains; independent studies of experienced devs on real codebases show far more modest —
sometimes even negative — net effects once review and rework are counted. Your own baseline is the only number that
matters.
| Metric | Trap | Better |
|---|---|---|
| Lines of code | AI generates lots of code — much of it wrong or unnecessary | Tasks completed, cycle time |
| Keystrokes/acceptance rate | Accepting suggestions ≠ correct code | Defect rate, rework rate |
| Surveys ("I feel faster") | Feels faster even when it isn't | Objective task time + quality |
| Speed alone | Faster code, worse quality = net loss | Speed AND quality AND cost together |
Measure time, quality and cost on the same tasks, before and after. That's a believable number.
Training that sticks
- Pair people on real tasks with an AI tool — learning by doing beats watching slides.
- Teach workflows, not features: "review this PR", "debug this test", "write this migration" — the actual task shapes.
- Share prompts and patterns as team assets — a shared prompt/skill library compounds (L2·01, L2·08).
- Internal docs and golden examples — this very site is the kind of artefact that works.
- Make room to fail cheaply — a safe sandbox where experimenting is expected, not risky.
Writing the policy
- What's allowed: approved tools/models, what data may be used (the gateway's approved list, L4·06).
- What's not: secrets, unreleased IP, personal data, anything regulated — named explicitly.
- Disclosure: when must AI-generated output be labelled or reviewed?
- Review duty: the human is accountable for AI output they ship. No "the AI did it".
- Keep it short and positive: "here's how to use it well" gets followed; "never do anything" gets ignored.
The human concerns
- Skill atrophy: if AI does all the writing, juniors never build the skill to review it. Deliberately keep "no-AI" practice and rotate who does the hard parts by hand.
- Job anxiety: be honest about what AI changes; hiding it breeds distrust. The durable skills are judgement, review and system design — exactly what this curriculum teaches.
- Uneven adoption: a few power users pull ahead; pair them with others rather than letting a gap ossify.
Try it (60 min)
- Take one real team task. Measure it without AI (time + quality), then with AI. Write the honest delta.
- Draft a 1-page AI usage policy for a team you know. Would people actually follow it?
- Sketch a 4-week adoption plan for that team: pilot group, training format, success metrics.
Key takeaways
- Adoption fails on people and measurement, not the tech.
- Baseline before AI, pilot small, train on workflows, scale with champions, measure honestly.
- "X% faster" claims are unreliable; measure time + quality + cost on real tasks yourself.
- Short, positive policy; the human who ships AI output stays accountable for it.