How our leadership team is learning to work together in AI

We're a company undergoing major AI transformation in how we work. I asked all our senior leaders to collaborate in Claude Code using a single shared AI workspace so they could better lead it.

I'm surprised by how many business leaders I know talk about AI with confidence, dare I say passion, but have only superficial experience with it. Leaders are doing their best by reading blogs, listening to podcasts, and scrolling through LinkedIn, but too few are getting their hands dirty hour to hour, minute to minute with AI. You can't lead an AI transformation based on secondhand knowledge. You've got to work with the technology firsthand to develop the judgment and intuition needed to successfully transform your organization.

We're beginning to see more executives admit this openly. Airbnb CEO Brian Chesky recently said he was intimidated by Claude Code and didn't know where to start, so he asked an Airbnb employee who was good with it to tutor him.

Most examples like that are individual leaders getting hands-on. Our goal goes a step further: for the leadership team, as a group, to build the intuition to lead AI transformation across the whole company. Our CEO, Adrián Ridner, has written that the manager is the lever for a team's AI adoption. The same is true one level up. A leadership team can't credibly ask the company to work this way if it doesn't.

I'm not a technologist, so I started with myself. In August, I switched from Claude Cowork, which I'd been using since it launched, to Claude Code, and began building a shared workspace for our leadership team, plus the skills, routines, and apps on top of it, mostly to learn what was possible and how to roll it out. For two months, only Adrián and I worked in it. I didn't bring in the rest of the team until a few weeks ago.

In my last Candova post, I described the Commons, the plain-text knowledge base we built so AI can find what it needs about our company quickly and efficiently. We left leadership material out of it on purpose. That was the right call, but it meant the work that steers everything else (our decisions, priorities, and who owns what) was the work our AI knew least about.

That work was also scattered across emails, chats, wikis, Excel, and Word documents. Each leader could get AI help on their own through our connectors to those systems, but nobody's AI could see across the leadership team.

So we made three changes.

1. Put the leadership team's work in one place AI can read

We call it the Leadership Commons. It's a shared folder of plain-text files. Like the company Commons, it tracks every edit, so we can see who changed what and undo anything. Twelve of us have access, and each of us can see everything. So far, it holds:

The decisions, topics, and to-dos we update every week live in a small database. Every night, AI copies them into the Leadership Commons, so the full history stays in one place.

Before anyone else wrote anything, we set one rule: if you don't mean to share it with the entire leadership team, it doesn't go in. Each of us keeps a private space for everything else.

2. Build skills, routines, and small apps that put it to work

On top of the Leadership Commons, I wrote instructions for the AI, so anyone who opens the folder gets an assistant that already knows how things are organized. Then I built three kinds of tools with Claude Code:

To help manage our leadership meeting, for example, we created routines to remind leaders when their update is due, turn the updates into a digest, draft the pre-read the day before, and draft the recap once the transcript is ready. The meeting itself runs on an app: a board where anyone can add a topic, update a to-do, or mark a decision made.

Connectors, which let AI reach our other systems, are also critical. A lot of this works because the AI can reach more than the files. Our setup gives our leaders, and most of our employees, secure read and write access to Microsoft 365 (email, calendars, Teams, and meeting transcripts) and full query access to our reporting data.

3. Ask every leader to do real work in it

Putting everything in one place and automating around it only gets you so far. If leaders still email their updates, the work lands in one inbox. The rest of the team sees it late, if at all, and the AI only sees a secondhand copy. So we asked every leader to do their leadership work in the shared workspace itself.

We used our planning offsite this year to get everyone working in it. I asked the team to submit their offsite homework in the Leadership Commons. Getting set up took three steps:

Before the offsite, we built apps on top of everyone's homework submissions. During the sessions, every leader used Claude to open those apps and work through real-time exercises: voting on, rating, grouping, and adding ideas. Meanwhile, AI grouped the room's ideas and logged open questions from each session's transcript. By the end of day one, we had a codified list of decisions and to-dos instead of a wall of sticky notes. Afterward, AI summarized the to-dos and kicked off the follow-up workflows. And because all the offsite instructions lived in the Leadership Commons, any leader could improve the process and materials, not just me.

When one leader files an update or builds a skill, every other leader gets the benefit.

We continue moving more of our leadership process into the Leadership Commons and building the AI workflows around it, but we're at an early stage. It's too soon to measure results, though a few things have already changed.

The part that's most exciting? One of our leaders built a new command for the whole team. They wanted their weekly update to turn into board items without retyping it, so they wrote /update-to-board and shared it with everyone. Now any leader can have AI propose board items from their update and approve each one. That's what we want more of: leaders improving the system for everyone.

More broadly, most of the team has jumped in and started figuring it out. Initially, Adrián and I made every change saved to the Leadership Commons: every filed update, logged decision, and fix to a skill. Since we opened it up to everyone, the other leaders have made about a third of the changes, and that's growing. A few leaders now do a lot of their leadership work there. Others use it for their weekly update and the offsite homework and not much else. Counting changes misses the leaders who mostly read and ask questions, so it's a rough measure. But it's a good proxy metric of leaders going deeper with AI to develop that intuition and judgment we discussed earlier.

The spread across our leaders won't surprise anyone who's read our post on AI readiness profiles. It shows up even among leaders who use AI every day. Every leader here already had access to AI. Building the habit of working in it together is harder, and it takes real work and coaching along the way. That's what we built Candova to do, and it's what we're still doing for ourselves.

It's not always clear where to work. We now have three places to work with AI: the company Commons, the Leadership Commons, and each leader's private space. My private space reads the other two as background, so that's where I start most days. But when I write something new, I still stop and ask where it belongs. A leadership note starts private and moves to the Leadership Commons once it's ready. An analysis could go in the company Commons, unless it touches something only leaders should see. If it's confusing for me, it's more confusing for everyone else. I flagged this in my first email asking the team to work this way, and we still haven't written down a clear rule.

It's not always clear when to build. As more of our work moves into Claude Code and the Leadership Commons, it's tempting to turn everything into an app. Our board replaced the Excel worksheets we used for decisions and to-dos, and that was the right call. Other cases are less obvious. Should a budget model live in a spreadsheet or an app? Should a draft that several people need to comment on live in Word or in the Leadership Commons? We don't have a rule for that yet.

The AI doesn't always have the latest picture. It can only work from what it can see, and our information doesn't all update at the same time. A leader can ask what's open and get yesterday's answer. A pre-read built the day before misses edits people make the morning of the meeting. We haven't solved this. For now, before I trust an answer, I check that I've pulled the latest from the Leadership Commons and when the routines last ran.

The biggest limit on what we've built is me. Every routine in this post runs as me, from my laptop. It reads my calendar, drafts in my Outlook, and waits for me to press send. If my laptop is asleep, the work waits. That's part of why so much of it still runs through me. As other leaders add their own skills and routines, those will run on their laptops too.

Across the company, our AI transformation is moving to always-on agent infrastructure: autonomous agents that run all the time on a server instead of on someone's laptop. We can use the same infrastructure for our leadership team. That could look like:

Perhaps our experience with always-on agent infrastructure will be our next post ;-)