A common question we get from the people we speak to about genAI is how companies are most effectively using it internally. Because these are general-purpose tools, it can be hard to know how to get the most out of them. We’ve written many articles about the use cases we’re seeing.
We’ve been curious to learn how the companies building the leading AI models are themselves using them for their own work. We recently spoke with Alex Albert, developer relations lead at Anthropic, about how the company uses its own tools, Claude.ai and Claude Code, internally. Here are highlights from that conversation, edited for length and clarity:
What are a few examples of how teams across Anthropic are using Claude in their day-to-day work?
The easiest thing you can start with is just our engineering and research. In those parts of the org, Claude Code has just taken off like a rocket ship. Folks are using it for working on Claude.ai, fixing PRs [pull requests]. Or if they’re doing research, they’re spinning up scripts, working within Jupyter Notebooks. It’s become kind of like our own junior developer on demand.
Shifting away from code, we see a lot of folks using just the general Claude.ai platform across the company [for] a variety of different things. So the typical ones are if you’re writing docs or you’re working on some sort of messaging or some sort of content creation, refinement, editing—that’s the most obvious use case for a simple chat back and forth with an LLM.
But what I think has gotten really interesting, especially over these past few months, is once you connect that chatbot ability with integrations. We use Google Drive, like Google Docs, Sheets, etc., internally. So folks are asking Claude to go search through all their docs and then synthesize things that they’ve read on a certain project and then pull out relevant notes. I use it all the time with meetings. I now transcribe all my Google Meet meetings—that automatically gets pulled into my Drive, and then at the end of the day, I can just ask Claude to summarize what I talked about, what my action items are, what have I done, what do I need to do.
As I’ve started to do a lot more product-management-type work for managing some of our new features, I use our Asana integration. I’m getting Claude to pull what the latest tasks are. I’m asking it to put in some specs—so if I just wrote a PRD [product requirements document], how can I translate that into to-dos for our engineers? We’re operating now in this expanded domain where Claude is more than just this back and forth that’s isolated to your one browser and it’s actually interacting with these systems that you use as a person in your own working day.
Are there any particularly creative or unexpected use cases you’ve seen emerge on teams within the company?
One thing that I think is really interesting, and this is somewhat in the coding domain but touches on a broader theme: We have some folks who are using Claude Code and they’re not engineers or developers, but it allows them to just do a lot more. One person on our [brand] team actually started shipping PRs and fixes to Claude.ai. Of course he’s not just directly pushing these right to the app or anything. They still need to go through a review from an engineer and all that, but that’s a fundamental unlock of which somebody with very limited technical background ran into a bug in Claude.ai and he’s like, ‘Hey, I want to fix this,’ just went into Claude Code and was able to actually put something up in review and get it fixed.
Also, people are using these things now for data analysis, which is a really big one. You’re a data scientist and you have access to BigQuery tables or logs and transcripts of some sort, and you want to compile all these sources and pull out the relevant insights and then make an HTML report from it. You can do that within Claude Code or you can do that within Claude.ai and have it write out to an Artifact. Then we can share those Artifacts internally as well. So you made this visualization, you have this really nice graph and you have these sliders, and it’s like a more dynamic interface than just presenting a PowerPoint slide deck. You take that, you can share that link internally, and now everybody can interact with this data modeling thing, which would have required an engineer in the past, but because Claude is able to write this code for you, because we can execute it for you, you can go that next step even if you’re non-technical.
How are use cases typically identified? Is it that you have guidelines and then anyone can experiment on their own?
Yes, that’s basically it. We have a good security team that makes sure that you’re not going to do anything crazy too, and the systems are all set up for that. We also explicitly encourage this type of iteration and prototyping across the company. So we have these hack weeks internally where we basically dedicate three, four days. People get into small groups and work on an idea of any type. There’s no objective. We have fun little prizes and things, but it’s not for a business reason, it’s just really to foster this culture of innovation and experimentation.
Our last hack week was back in October. And the thing at the time that captured all the rage in the company was MCP [model context protocol, a way to connect LLMs to data sources and tools.] This was before MCP launched, but the hack week was absolutely dominated by all these folks who were building their own little websites where they could plug in MCP servers that they had Claude create. People were controlling robot arms, people were controlling a 3D printer. People were exploring some of these initial integration ideas before we had even built them out into Claude.ai. That was a really telling moment for us to see, ‘Oh, wow, there’s something to MCP.’
