Most dashboards measure what a team produces. Almost none measure how fast it gets better — and the two come apart quickly. A team can hit every deadline for a year and be no more capable in December than it was in January.
Fig. 1 — The first line is the one most organisations measure. The second is the one that decides whether the team is worth more next year than it is today.
Dig into the research on high-performing teams and the same small set of habits comes up every time. None of them are complicated, and every one can be learned. What gets me is that each one is a thing that happens between two people. Which is exactly why AI changes them — helping with some, and undoing others without anyone noticing. Which way a habit goes follows a rule: the ones that ran on effort — trying, critiquing, funding — get cheaper, and the ones that ran on evidence — the visible gap, the reported blocker, the customer in view — lose the evidence they ran on.
Make it normal to try things
Between the one who tries ↔ the room that reacts to a miss
The teams I have watched improve fastest run far more experiments than the rest, and the difference starts with the cost of being wrong. Asking people to be bold while punishing every miss gets you the behaviour you punish. So name a failure budget out loud, give half-finished work somewhere to live, and let the most senior person describe their own experiment that went nowhere. On one team we wrote a rule down: nobody writes more than a page before showing it to someone.
The cost of trying collapses. What took a week takes an afternoon, and three options are cheaper than one was. The constraint moves to deciding what’s worth building and reading the results honestly. The failure mode is a hundred experiments nobody learns anything from, so keep a plain record of what was tried and what it taught.
Say what you don’t know, first
Between the senior voice ↔ everyone taking the cue
Curiosity travels downward. If the most senior person in the room asks a real question, admits a gap, or changes their mind in public, everyone else gets permission. The opposite also holds, and it’s expensive: a team that thinks the leader has all the answers stops offering theirs.
AI makes gaps invisible. Someone can hold a fluent conversation on a subject they don’t understand, and nobody can tell, including them.
Ask what’s blocking them
Between the one who’s stuck ↔ the one who asks
The question most leaders avoid is the one that helps most: what are you stuck on. It works when it comes in a rhythm people can rely on, and when it arrives somewhere they aren’t being assessed.
Years ago in London, a client sponsor suggested we hold our one-to-ones in one of the Wren churches tucked between the office towers, run as cafés now. Coffee under a 17th-century ceiling once a month, lunch now and then. We heard about frustrations months before they hardened into positions, and I’d credit the room for more of that than anything I brought to it.
The evidence disappears. People are less likely to report being stuck, because they were unstuck at 23:00 by a model and nobody saw it happen. The work still ships and the team stops seeing where it was hard. So ask differently: what did you have to work around this week? What did you ask a model that you’d have asked one of us a year ago? Those answers are where your process is broken.
Stay in the work
Between the leader ↔ the work itself
Standard advice sends leaders up to strategy and out of the weeds. The evidence points the other way. The leaders I have watched build the strongest teams are more in the work, not less; managers who supervise from a distance lose visibility and spend their days reacting to problems they never saw coming.
The line between involvement and micromanagement is simple enough. Working alongside someone builds their capacity; hovering takes it away.
AI makes distance comfortable. Summaries, digests, a dashboard, and you can sound well briefed on work you never touched. Be suspicious of that feeling. Use the model to prep for the room, then get in the room.
Make feedback feel like help
Between the note-giver ↔ the one receiving it
Feedback fails to improve performance in about a third of cases, and sometimes makes things worse. Usually because it lands as a verdict. What works is frequency and low tension: short, regular, specific, well before anything is final.
Most of what I know here I picked up from my wife, who trained as an ICF coach. In London we built a peer-to-peer coaching framework and taught the basics across the team, so people could frame their own thinking and hold a colleague through a stuck moment. We rehearsed these with our team & actors, which sounds odd and worked.
A model will critique your draft twenty times without a flicker of impatience, and that is useful. Let it take the mechanical pass. Bring the human pass to the questions it can’t answer: is this the right problem, will this land with that stakeholder, does this sound like us. The risk is outsourcing all critique to the machine and never practising the harder conversation — a skill that fades when unused.
Back growth that costs you
Between what you fund ↔ what people conclude
The strongest teams support growth that doesn’t obviously pay them back. Side projects, training that points somewhere else, people who leave well. Those people come back, or they send someone better.
This matters more now, because the anxiety in the room is about craft. People are watching to see whether AI is being used to make their work bigger or to make them cheaper, and they will decide from what gets funded, however good the town-hall speech. Back the learning that makes someone more valuable, including elsewhere. It’s the clearest signal you can send that you mean to keep them.
Say why it matters
Between the team ↔ the person the work is for
Leaders of the best teams are markedly better at connecting the work to who it’s for. That is proximity: getting the team near the people the service serves, often enough that the abstraction wears off.
AI arrives wrapped in an efficiency story, and if cost per task is the only story told, motivation follows it down. Say what the capacity is for. Faster service for people who are currently waiting, harder problems, work the team couldn’t attempt before. Nobody ever did their best work to save the company money.
Fig. 2 — Four of the seven get harder when AI arrives, because the habit ran on evidence AI clears away.
This is a behaviour decision, which means you can check it this week: listen to your team.
- Someone saying they’re stuck.
- Someone showing the rough version.
- Someone asking the question they were scared would make them look slow.
- No one stuck. No rough drafts. Everyone equally informed.
- Silence — and that catches up with you.
Fig. 3 — The test takes one week and no budget. You are listening for the sound of people learning in the open.
So pick one of the seven and put it in your next one-to-one this week. If you want a number for the second line, count the rough things shown before they were finished this sprint; it is the cheapest leading indicator I know. If people start naming what they’re stuck on and showing you the rough version, the team is compounding. If the room stays smooth and quiet, lean in and listen.
Sources: Kluger & DeNisi, “The Effects of Feedback Interventions on Performance” (1996) — the meta-analysis behind habit five: across 607 studies feedback raised performance on average, yet reduced it in over a third of cases. On psychological safety as the ground under habits one to three: Amy Edmondson (1999) and Google’s Project Aristotle.