On many engineering teams using AI today, a few engineers have found a way of working with it that makes them much faster, and the rest of the team is still finding its feet. A CEO hears about the fast ones and expects the whole team to look like them before long, but what one engineer has learned stays in their head unless the team has a way to share it. Closing that gap, so what one engineer learns reaches the whole team, is Learning, the fourth of the Four Handoffs, and it’s where a lot of what AI could do for a company leaks away.
People learn it from the people beside them
Two rollouts show how use of AI spreads. When Microsoft studied its own early 2026 rollout of coding agents, an engineer whose manager used GitHub Copilot CLI had 82% higher odds of trying it, and the link was stronger still when much of their wider peer group used it (Microsoft, 2026). Docusign reported the same pattern in its engineering teams: giving every engineer access stalled at 49% active use, a six-person central team raised it to 65%, and once working engineers inside each team became the go-to peer mentors, active use reached 93% within two months (Docusign Engineering).
Spreading it takes leadership inside the team
An engineer who is really good with AI is often not the one who teaches it best, so turning one person’s skill into the team’s takes leadership inside the team. The aim is a culture of sharing, because a rising tide lifts all boats. Sharing catches on when a team already talks openly about its work, and it lasts when there’s a habit the team keeps every week and a page the team can come back to.
Someone answers for every agent
Every agent on a team needs a person who answers for its work. Research from BCG, published in Harvard Business Review in May 2026, found that when AI was described as an employee, managers’ sense of individual accountability for errors dropped by 9 percentage points, and that’s the risk to guard against. Microsoft’s 2026 Work Trend Index puts agents on a scale from assistant to teammate and says people stay involved by setting direction and taking responsibility for how the output is used. So the engineer who asks an agent for a change owns it as fully as code they wrote by hand, no agent approves its own work, and the rules agents work by change only when your CTO, or the owner they name, approves.
The Weekly Replay
The habit that spreads it is short. Once a week for thirty minutes, one engineer replays a real session with an agent, including where it went wrong and how they caught it, and the team talks about the judgment involved more than the tool. Most weeks it ends with one change to the instructions the whole team’s agents work from, reviewed like any other code and approved by your CTO or the owner they name. It’s our team-level form of what Thoughtworks calls a feedback flywheel, and it means the practice outlives the one power user, because what they figured out now lives in everyone’s agents.
What a CEO should see
We measure all of this by team, never by person: how long work takes to reach customers, how long work waits on review, how often a change fails in production, and how often agents repeat a mistake the team has already fixed, a number that should fall as each lesson spreads. Usage tells us where to look, and it never becomes a leaderboard. Set beside what you spend on AI, those numbers show whether the gains are reaching the whole team’s delivery or staying stuck.
If one or two of your engineers are flying with AI and the rest of the team hasn’t caught up, I want to talk to you.