Adoption is a curve. Every pilot follows it.
By the end of this lesson, you'll be able to read the adoption curve, recognize the two dynamics unique to AI tool rollouts, and make the right call on when to intervene and when to let it run.
From supervised to agentic: a predictable progression
Every Claude Code adoption follows a three-phase curve. The phases don’t vary much by client. The duration and depth of each phase vary a lot.
Developers run Claude with manual approval for every action. They’re building trust in the output and calibrating their prompting style. Velocity is often slower than without Claude. This is expected. Don’t intervene on velocity in Week 1.
Developers begin approving more actions in advance and running Claude on larger task chunks. Plan mode usage increases. Some begin exploring auto-accept mode on low-risk tasks. This is where CLAUDE.md quality starts to matter.
Developers identify specific task types where Claude’s output is reliable and hand those off with minimal supervision. PR cycle time improvement becomes visible here.
Usage is routine, not experimental. The team has developed informal norms about what to delegate, how to supervise, and how to brief Claude for new task types. This is the target state. It can’t be forced earlier.
Two dynamics that don’t appear in ordinary tool rollouts
Treat Claude Code like a SaaS rollout and you’ll misread the early data and intervene at the wrong moments.
Developers are learning a new interaction model, not just a new tool. The prompting loop, the supervision decisions, the judgment about what to delegate: these take time to calibrate. Usage will look low and output will look inconsistent. This is not a signal to intervene with training. It’s a signal that calibration is happening.
After the Week 1 dip, many developers find a comfortable usage level and stop experimenting. They’ve learned one or two delegation patterns that work and they stick to them. Without active champion intervention or a structured Week 2/3 session (Code 201, observability depth), plateau can look like adoption but the ceiling is much lower than it should be. This is when champion enabling matters.
A developer is stuck. The question is whether to help now or wait.
The wrong intervention at the wrong moment sends adoption backwards. There is a simple test.
The test is whether the developer is stuck on a Claude Code problem or a workflow design problem. If they can’t get Claude to do what they want (hallucinated output, prompt failures, context issues), that’s a technical intervention — help now. If they’re doing the same tasks they always did and Claude is just a faster autocomplete, that’s a workflow design problem — the delegation pattern needs rethinking, not the prompting technique.
Week 1 interventions: technical calibration, CLAUDE.md setup, first CLAUDE.md commit. Don’t push for velocity.
Week 2 interventions: delegation pattern review, plan mode encouragement, Code 201 session. Watch for plateau.
Week 3 interventions: auto-accept mode for defined task classes, hook configuration, agentic task structuring. Look for the teams whose usage is still climbing vs. those who plateaued.
Week 4: document the usage patterns. What worked? What broke? What should the team’s CLAUDE.md say that it doesn’t say yet?
A developer who has stopped experimenting is not a success story. Plateau is not adoption.
The Helios Software Day 16
Helios Software. Day 16 of a 30-day pilot. 10 developers. Java-to-Kotlin migration use case. The data is starting to come in. Three decisions, up to 6 points.
PR cycle time is flat compared to baseline. Usage data shows developers averaging 45 minutes per day in Claude Code. Prompting patterns haven’t changed since Week 1.
What does this tell you?
You’ve identified a plateau. What’s the right move?
Which intervention best breaks the plateau at Day 16?
After the Code 201 session, one developer (Priya) has made rapid progress. Her PR cycle time is down 40% from baseline. The rest of the team is still at plateau.
How do you use Priya’s progress?
You can read the adoption curve. Now build the team behind it.
Adoption is a design problem, not an enthusiasm problem. Plateau is predictable and addressable. The three-phase curve gives you the language to describe what’s happening and the intervention map to respond.