# Show Me Your AI with Danbot Gilbot — How to Roll Out Claude to 1,000 People *Guest: Poppy Bryant (Director of Operations)* --- **Dan:** Welcome to Show Me Your AI with Danbot Gilbot. Today we are talking about how do you roll out Claude to a thousand people? There is so much that goes on behind the scenes to actually get stuff done. I'm here with Poppy Bryant, our Director of Operations, to talk us through this — not a generic vision, but the actual mechanics and breakdown of how we really did this. Let's get into it. Welcome. **Poppy:** Thank you. ## Step 1: Decide on your infrastructure **Dan:** What's the first step? **Poppy:** The first step is that you've got to decide on how you'll set up your Cowork infrastructure. That includes your frameworks and ways of working. You need to think about your skills workflow. And obviously you need to sort out your Anthropic contract if you're going with Claude. And you've got to think about all of the communications that you're going to have around it and all of the people you need to include. **Dan:** Yeah, this is the more logistical stuff that you need to set up. We're talking more today about how you actually roll it out. ## Step 2: Start with hand-raisers and product champions **Dan:** Talk us through the less mechanical, onto number two. How do we initiate this? **Poppy:** You need to start off with a smaller group of hand-raisers and product champions — and they're probably people who have already used it. They're often technical people, or maybe you do have some practice leads or junior people. If you get them working it before you open it up, you can make sure that you've really tested, you're thinking about the architecture and how you want people using it. **Dan:** And it pressure-tests the framework. **Poppy:** We knew who some of these people were because it allowed people to set up licenses on different platforms, so it's quite easy for us to identify. Each company is going to be different. **Dan:** But what you're saying is, because it's a new rollout, you can't necessarily filter by the people that are already using the tools. You have to try and discover the people that are actively doing it, maybe on weekends and evenings, which I think we have with some people, right? **Poppy:** Yeah, exactly. And because we have a lot of Slack channels, there are people who are shouting about it — the people who are posting about all of the cool new releases. They're the people you need to tap into. **Dan:** Not to get too meta, but you can actually just ask Claude, "Look at all of my Slack channels and tell me who's talking about and using AI the most in public channels." There were some instances where they had probably gone too far, or too hobbyist. There were some open Claude instances which we didn't castigate people for, but it created a security risk. **Poppy:** Yeah, but you have an amnesty and you can help them bring out the really great ideas they have, and turn it into an enterprise-level rollout. So we caught some lobsters, said we love this — let's try and turn this into enterprise infrastructure as part of our rollout. ## Step 3: Launch, even if it's not perfect **Dan:** Excellent. So we get some people, and then? **Poppy:** You've got to launch it. Even if it's not perfect, you've just got to get it done. And you start by showing off the results from that smaller pilot group. **Dan:** When you say launch, this is launch for the whole company, but with some structure in place already. **Poppy:** Yeah, exactly. You've got it structured with the architecture that you designed to begin with, but then tweaked based on the testing from that smaller group. And just before you launch, you show off some of our examples. You need to hype it up — people want to get logged in as soon as they hear it's being released. **Dan:** And as part of that release, we wanted to launch with something super useful. One of the biggest complaints we had is people hate making presentations. Every single day someone would say, "Has someone built an agent to generate my presentations yet?" This is maybe a marketing-agency-specific thing, so it might not be relevant for everyone, but it was that specific to us — people were saying this is their least favorite job in the world. So we basically found the biggest pain point and then presented a solution for that. Plus it's a universal skill or agent, not just a very specific thing that works for one person. So that's what you anchored on and said, hey, we can solve your least favorite thing to do. **Poppy:** Yeah, give it a go. **Dan:** I like that. So we built that. We'll showcase that one in particular just to get people excited about it. ## Step 4: Training — a "doing" training, not a "watching" one **Dan:** What's next? **Poppy:** To get people in, you do have to do some training — a bit mechanical again, but you need to know who's running each of the sessions, what's being covered in those sessions, and most of all what people need to do once they've left that session. **Dan:** We've actually shared a bunch of our training material in a separate video that we'll link to. What were the key takeaways or key parameters that you put into the training rollout? **Poppy:** Some of the key parts we wanted to include are the basics, the fundamentals of how everything works. But the main one is you don't want to let anyone leave the sessions having not logged in or not used it for the first time. **Dan:** This is a "doing" training, not a "watching" training. **Poppy:** 100%. **Dan:** So each person logs in and is told to do what — run an actual skill, like try the presentation skill, or any skill? **Poppy:** Yeah, exactly. Step one, try the presentation skill or a super easy skill. And once they've done that for the first time, we also did a bit of a workshop where they got together in their teams and were able to come up with ideas to build new skills themselves. **Dan:** Right, so not conventional learning as such — more like a workshop. ## Step 5: Set the skills architecture and guardrails **Dan:** What's next? **Poppy:** We want to make sure that people have the right parameters and guidance on how they use it moving forward. During that training, we wanted to make sure people understood the flow and the architecture that we've set up. So when they build or create their skills, they know it's going to go to someone to check it, make sure it's in line with best practice, and then it gets approved so everyone can use it. And as they use those skills, other people can feed back on them as well. We have this iteration loop of skills always improving. **Dan:** Yeah, that's our skills architecture. It's a system where it's not just each person creating their own skills — they're not pitching tents, they're contributing to a palace, which was a key part of how we improve organizationally. So we were saying: this is how to actually submit a new skill, this is how to discover existing skills, this is how to suggest updates. That's what they were being trained on and shown in the actual interface, so they don't all go off making their own things and vibe-coding apps that have been built elsewhere. So we are live — we've given everyone licenses, they're all in the right architecture, they all know what a skill is, they've all used a skill, they all know one that works, and now they've got access to the repository. Each person now knows this is what Claude is, this is where I log in, this is how I use it, here are some of the parameters and guardrails and safety rails. Clearly that's not quite enough to get what we'd call a high level of AI nativity, or "rollout complete." That's the beginning, not the end. So what's next? ## Step 6: The Trio model **Poppy:** There's a bunch of things you need to do to get something to really land at an organization. But one of the most innovative things was our Trio model that we set up and made people aware of. You have someone day-to-day in the work — our product champions, who build the skills. They're the people within the teams who know the workflow and spot the opportunities, and skills enable these people to just create these workflows. But sometimes they might need that extra technical lift from a tech partner. So if the skill doesn't quite work, or there's not the connector that's needed, or there are some logic gaps, the tech partner can step in and advise on different tooling or diagnose the issues the person might be having. **Dan:** For us the tech partner is a classic software engineer that really understands limits of scalability and what's going to break. Whereas the product champion, in many instances for us, was even some of our graduate, early-career people, because they just have the time to go deep and learn all of the materials in and out of the platform. The third being the practice or department lead. These are the people who really know what good looks like for the work itself and what outputs need to happen, but they also help drive adoption and make sure those skills are generated at a great pace and help identify those opportunities. **Poppy:** For instance, if it was an SEO specialist, they'd be the person that actually does SEO work and monitors whether it's a good piece of work. It's those three people that work together. **Dan:** One of the reasons we went down that Trio model was pragmatic — there's not that much AI-native talent. These tools are really new, and having all of those skills combined in one person is quite rare. So we were working around that to say, let's put these three people together. But I imagine six to twelve months forward, this is something we'd expect one person could lead. ## Step 7: Track engagement data **Dan:** With all of those different activities, we need to keep an eye on engagement data. You want to know how things are landing, who's using it, who's not using it and why — and ring all of those people up and discover it. Because Cowork is a relatively new product, there's not a huge amount of functionality to begin with on tracking and analytics, so we built quite a lot ourselves to begin with — though Anthropic is releasing more and more every week. It's not just who's using it; you want to know how they're using it and whether they're using it efficiently to get super-productive outputs. Talk us through what we're tracking. **Poppy:** We have active days within the month, the number of chat messages, number of Cowork sessions, Cowork messages, the number of connectors used, agents used, agents created, agents updated, and then files, artifacts created, and code lines as well. **Dan:** The interesting bit here that maybe other people haven't thought about is the created and updated, because that's how you get the ongoing improvement of all of your agents. **Poppy:** All of these agent stats and metrics mean we can spot patterns. We're actually building out a heat map of different usage across the org, so we can identify where we might need to step in and see what the barrier to entry is — or see who's using it really well and take their examples and give them to other teams. **Dan:** Spotted some clusters, but we want a visual so we can see that by region, by practice, by leader, whatever it might be, then just hone in on those. **Poppy:** Exactly. Then we can continue to build this really vibrant ecosystem of skill creation, skill usage, improvement, iteration — and your point around everyone having the skills they need in a year's time. We'll get there. ## Summary **Dan:** Love that. Thank you, Poppy. This was super helpful. Just to sum up: 1. Decide on the infrastructure. 2. Get champions, build some skills so people see instantly how good it is when you roll it out. 3. As part of training — I don't even want to call it training, I want to call it onboarding — people use it, showing how good it is. 4. The Trio model, which may become a solo role at some point, but for now combines the different skill sets that don't generally exist in one person. 5. Capture all of the data and use it to hone in on the winners and celebrate them, and find and execute the losers. How do you actually do this in practice? Get on with it. That's the Show Me Your AI. Real enterprise AI in a real company, no theory. Signing off, Danbot Gilbot. See you tomorrow.