From pilot to scale

The framework

Pilot → Champion → Scale → Govern

Each stage has a clear gate. Don't skip stages — most failed AI rollouts skipped from pilot to scale without the champion stage.

1 — Pilot (weeks 0–8)

Pick one agent. The right first agent has three traits:

• Real pain — not "this would be cool"; "this wastes 4 hours of every Friday."
• Bounded scope — one team, one workflow, not "all of HCM."
• Measurable outcome — cycle time, accuracy, volume — something you can put a number on.

Pick the cohort.

• 10–30 people maximum.
• At least one skeptic (their objections sharpen the rollout).
• At least one champion (their advocacy carries the rollout).
• Mix of seniors and juniors.

Run the pilot.

• Train: 30 minutes is enough for an embedded agent.
• Monitor: weekly pulse check with the cohort.
• Capture every issue, however small.

2 — Champion (weeks 8–16)

The pilot ends; the champion begins. Three things happen:

• Tell the story, in numbers and quotes. ("Cycle time down 40%; here's what Maria says.")
• Train the next cohort using the pilot's lessons.
• Fix the obvious gaps — the issues from pilot that didn't block but did annoy.

This is where most failures happen. Teams declare victory after pilot and skip the championing — so when the wider rollout hits a bump, there's no constituency to push through it.

3 — Scale (weeks 16+)

Now expand to the full population. The champion phase has built the case and the materials. Watch:

• Adoption curve — if usage stalls in week 4 of broad rollout, something's off.
• Override rate — if users override the agent's suggestions >50% of the time, the agent isn't fit-for-purpose.
• Complaints — listen to what people complain about, not just that they complain.

4 — Govern (ongoing)

Once an agent is in production:

• Quarterly accuracy review — does it still meet the bar that justified the rollout?
• Quarterly permissions diff — has scope crept? Reset.
• Annual fairness audit for any agent involved in people decisions (hiring, promotion, comp).
• New-agent intake process — every new agent goes through the same Pilot → Champion → Scale → Govern path.

The four most common pitfalls

1. Picking the wrong first agent — going for novelty over pain.
2. No measurement — "we'll just see if people like it." You won't know.
3. Skipping champion stage — declaring victory at pilot, then watching the rollout stall.
4. Over-trusting the agent — auto-approving where a human should still be in the loop, eroding trust the first time something goes wrong.

Avoid these and the rest is mostly execution.