Inside a Claude Implementation

No newsletter last week. My ten-week streak is broken and I’m not happy about it. Hopefully somebody noticed — if nobody did, that’s a bigger problem than the missed week.

I was on-site with a company to get their team set up on Claude properly — creating accounts, installing Claude on every machine, integrating third-party apps via connectors or APIs, training on practical use cases. In other words, I was busy with a “Claude implementation” (a term I made up), and it pulled me out of the writing chair for a few days.

(If you’re still using Claude in your browser, download the desktop app. I’ve lost count of how many conversations I’ve had with people asking Claude questions who are one click away from a better experience with more features. The desktop app is free, it takes two minutes to install, and it makes a huge difference.)

Most Applied AI readers have never seen what a Claude implementation actually consists of. The four things I build for a client are the same four things you should be building for yourself inside your own org.

A “Claude implementation” isn’t a training session (although it does include workshops). It’s a set of concrete systems the organization keeps using after the initial work is done. When I work with a team, the deliverables are the same every time.

Workflow map. A written audit of how the work gets done today. Where the bottlenecks are, where people are wasting time, where existing software is already half-solving a problem. Most of what surfaces in a workflow map isn’t an AI problem. It can be as small as a Gmail filter, a template, or a process that nobody has touched in years. Claude gets better results when it’s applied to a workflow that has been examined first.

Team fluency. Installing Claude, setting up accounts, activating the right connectors, and walking through real use cases. By the end of an implementation, everyone on the team should be using Claude on their own work.

If you’re feeling lost, Anthropic Academy is a great starting point. I’d recommend Claude 101 and Intro to Cowork courses. Use Claude every day for a week on live work. Fluency comes from regular practice.

Company brain. A structured enterprise project with sub-projects by function. Standard operating procedures, client policies, key contacts, anything that people currently have to dig around for. Searchable in plain English. The goal is that Claude has context on how the business actually runs. Anthropic calls this “enterprise search” — like Google but all of the search results are based on internal company docs.

If you’re not an organization owner, you can replicate a version of this by setting up a project with context on your recurring work. Your role, your clients, your writing style, the documents you return to. It takes a few hours to build and it changes how every future conversation starts.

Configured systems. Role-specific Claude projects for recurring work, plus — for some clients — custom tools built on top of the API. The projects are something any professional can build for themselves. Claude Code makes building custom tools possible, but that’s a conversation for another day.

These four deliverables are the basics. Nothing clever, nothing novel. But they are what separates the people who get serious leverage out of Claude from those who are stuck in beginner mode.

The reason to do them in order is that each one makes the next one easier. A workflow map tells you where to focus when you’re building fluency. Fluency tells you what context Claude actually needs in your company brain. The company brain is what makes your configured projects useful. Skip a step and the later steps don’t land the same way.

Notice this week when you reach for Claude and when you don’t. When you’re not using Claude for work, take a moment to ask yourself why. The barrier between human work and AI work is the outline of your workflow map. If you’re having trouble visualizing it, sketch it out and upload a photo to Claude. Or, if it’s easier to talk through, just dictate it to Claude.

If you’re reading this without a paid Claude subscription, get one. Pro is $20 a month and it’s the single best productivity investment available to knowledge workers right now.

If you don’t want to pay for it, I will. At the bottom of this email is a referral link unique to you. Share it with people who should be reading Applied AI and get rewarded for it:

Waymo's cars are filling potholes now. Earlier this month, Waymo announced that its self-driving fleet is automatically flagging road damage and piping the data into Waze, where cities can pull it for free. Five metros live, around 500 potholes identified in the first rollout. Nobody built self-driving cars to be a civic infrastructure sensor network, but here we are. Turns out that a fleet of vehicles with perception systems driving the same streets every day is the best pothole detector a city could ask for.

Claude Design launched Friday. Anthropic released a new product that lets you build prototypes, decks, one-pagers, and mockups by describing what you want. Powered by the new Opus 4.7 model. Available in research preview to Pro, Max, Team, and Enterprise subscribers. Figma stock fell 7% the day it launched, which tells you how seriously the design world is taking it.

Allbirds is an AI company now. The wool sneaker brand that once defined Silicon Valley business casual announced last Wednesday that it’s pivoting to AI compute infrastructure and renaming itself NewBird AI. The stock jumped 582% in a single day before giving most of it back the next morning. Allbirds sold its actual footwear business to American Exchange Group for $39 million last month. I’m a little sad because I liked Allbirds (even after the shoes became a Silicon Valley meme), but it’s one of several DTC brands that should have never gone public.