Everyone's talking about AI. Solopreneurs are showing off impressive demos. Consultants are publishing frameworks. LinkedIn is drowning in hot takes. But nobody is showing you how to actually make AI useful at a mid to large size company. On this page I will not add more theory. If you need convincing that AI matters, this page isn't the place. We're past "should we?" and deep into "how?" What follows is a field report. I'm Dan Gilbert, CEO of Brainlabs, a media agency of over a thousand people across NYC, London, LA, Dallas, Boca Raton, Toronto, Buenos Aires, Singapore, and Sydney. For the past six months, we've been rebuilding this company as AI-native. I'm sharing the infrastructure, the architecture, the technology choices, and the governance model so that you're not left to figure everything out by yourself.

## The Core Problem: Inspiration vs. Perspiration

There's a useful idea from David Deutsch. He separates knowledge creation into two layers. Inspiration is the creative part: conjecture, direction-setting, judgement calls. Perspiration is the execution of those judgements. It's mechanical. You can absolutely reduce it to a flowchart. Anything that can be reduced to rules can be automated. AI is extraordinarily good at perspiration. It cannot set direction. But once you've said "go north," it can get you there faster and more thoroughly than any human can. Most companies are structured backwards. Their best people spend 80% of their time on perspiration work: formatting, researching, checking, iterating, scheduling, coordinating, chasing, copy-pasting. The remaining 20% goes to actual thinking. AI flips that ratio. Your best people should spend 80% on inspiration and delegate the perspiration to machines. Tobi Lütke, the CEO of Shopify, now requires teams to demonstrate why AI cannot do the job before they're allowed to request new headcount. The burden of proof has flipped.

## A Quick Guide to the Jargon

Before we go further, some definitions. • A prompt is a one-off instruction you give an AI. You type it, you get a response, it's gone. • A skill is a prompt that's been refined, tested, and saved so anyone in the company can reuse it. Prompt = ad hoc. Skill = codified. • An agent is AI that can take actions, not just answer questions. I like to think of agents as junior employees. You give them clear instructions and they go carry out the work. • A wrapper is the software layer that sits around the AI model and gives it memory, file access, and structure. The wrapper is arguably more important than the model itself.

## Step One: Choose Your Tech Stack

If you let different teams build AI systems independently, you get fragmentation. Team A uses one model. Team B uses another. Finance builds automations in Zapier. Marketing is somewhere nobody knows about. Every team is "doing AI." Nobody is building anything that compounds. The real cost is threefold. First, knowledge leaves when people leave. Second, nobody learns from anyone else's work. Third, context is shattered. Our first step in becoming AI-native was to choose one organizational platform for where all of our AI, all of our agents, and all of our institutional knowledge would live.

## Why Notion?

We evaluated a lot of options. Notion was the best for us. 1. Work already lives there. Notion is already a best-in-market solution for project management, task tracking, meeting notes, the company wiki. 2. Notion is LLM-agnostic. A company can't afford to bet everything on one model. With Notion as the wrapper, if Gemini is the right LLM today and something else is better tomorrow, we can switch. 3. Notion already has enterprise features we needed: permissions, version history, audit logs, integrations. 4. Notion is accessible. Everyone in the company can browse the skills database and read in plain English what every agent actually does. If you can't explain it in a Notion page, you don't understand it well enough to automate it. 5. Notion also serves as our context layer. The principal difference between a good AI result and a bad one is context. Ask an AI to "write a competitive analysis" and you'll get generic slop. Ask the same AI while it has access to your client database, your methodology, your past examples — you'll get something remarkably useful. The model matters less than what the model knows.

## Why Claude?

We use different models for different tasks. But when it came to choosing the primary execution layer, we chose Claude. The reason is less about the model and more about the tooling Anthropic built around it. If you've used ChatGPT in a browser, you know the experience: you type, you get a response, and then you start over. No memory. No file access. No persistent instructions. That's fine for one-off questions, but it's completely useless for real work. Anthropic ships two versions of this. Claude Cowork is the one most people in our company use. It looks and feels like a workspace. Claude Code is the command-line version for our dev teams. As of writing, Anthropic is the only company that has built wrapper tooling at this level. Other companies have built models. Anthropic built a system for models to work in the real world. That's the difference, and it's not close.

## What Daily Work Actually Looks Like

Take an SEO audit. Before AI, that's a multi-step process: set up a crawl, export the data, build a workbook, categorize the issues, score them by priority, create charts, write up findings. Several hours of perspiration work for every audit. Now, the analyst opens Claude Cowork and tells it what they need. In plain English. Claude understands what the analyst is trying to do and automatically identifies the right skill from our library: the technical SEO site audit skill, built and refined by our best practitioners. Claude pulls that skill, follows every step, asks clarifying questions where judgement is required, and produces the deliverable. The analyst reviews, applies expertise where the AI got something wrong, and ships it. The skill doesn't live on one person's computer. If someone finds a better way, they suggest an update. If it generalizes, everyone gets the better version. The next audit is better than the last one, automatically. Person opens Claude. Claude calls a skill from Notion. Skill runs. Person reviews and applies judgement. Output ships.

## Management Teams: The Trio Model

Each department has a trio of leaders responsible for their skills. The first is a domain expert: someone who works in that practice and really understands what good output looks like. The second is a tools specialist: someone who's gone deep on Notion and Anthropic's training. The third is a technology specialist from our engineering team for scale problems. Below the trio is a community of users. They run the skills every day in Claude Cowork. They flag when something is broken or it can be improved. The trio reviews and asks one question: does this generalize across users, or is it specific to one situation? If it generalizes, everyone gets the better version immediately. We think of this as building a palace, not pitching tents. A tent is an individual automation on someone's laptop that dies when they leave. A palace is shared infrastructure that gets better with every person who contributes to it.

## Token Economics

Every action in Claude uses tokens. If you're going to run this across a thousand people, you need to understand what drives cost. Most people assume the prompt is the expensive part. It usually isn't. The cost comes from context: everything Cowork pulls in to do the job. Tell it to search all your Slack channels for anything about a project and it'll read hundreds of messages. Tell it to search one specific channel from the past two weeks and it'll get there faster, use a fraction of the tokens, and produce a better result. The habits that keep token use down are exactly the same habits that produce better output. Be specific about where to look. Use a dedicated folder, not your entire shared drive. Ask for the plan before it starts acting. Focused use and good use turn out to be the same thing.

## Are You Replacing Staff?

Some rudimentary math: if AI can replace 20% of tasks and I have a team of 5, then I can either do 20% more work or remove one person. Given the choice, I'm growing 20% without adding headcount. Every business owner I know thinks the same way. Every major technology shift in history created more jobs than it destroyed. The printing press. The industrial revolution. The internet. The Doom is completely overblown. But the doom might be useful for one thing: scaring people into actually learning the tools. AI won't replace you. Someone who uses AI well might. What you've seen in this article is not about replacing people. It's about removing the work that isn't really work and giving people back their time for the work that actually matters: thinking, creating, making judgement calls, solving problems that haven't been solved before.

## How to Think About the Phases

Phase 1: Organise your data layer. Pick a single system of record. If your knowledge is scattered across drives, wikis, and people's heads, AI has nothing to work with. Don't skip it. Phase 2: Select your tooling and build your first skills. Build for the highest-volume, most repetitive work first. Get early adopters using them daily. Measure time saved. Phase 3: Roll out training and scale. Hands-on sessions where people build skills for their own work. Establish governance. Phase 4: Integrate and automate. Connect your AI layer to your core systems. Work starts flowing through the system with less human intervention at the routing level. This isn't a project with a go-live date. It's a permanent shift in how your company works. The companies that move first will compound their advantage every month.

You can do this. Your stack might be different. Your industry might require different routing logic. But the pattern applies: one platform, one set of skills, one governance model, one feedback loop. Inspiration stays human. Perspiration gets automated. My advice is simple. Sign up for Notion. Learn Claude. The only way to get good at AI is to use it. Just get on with it.