I set up our email lists recently. Then the email sequences that run off them, then the Google Analytics and Tag Manager tracking that follows a visitor through our site. I didn't brief a developer for any of it. I described each job in plain English, and Claude carried it out inside the tools themselves.
That's the part worth noticing. Everyone knows by now that an AI model can write an email. Fewer people have seen it do the setup work that used to sit in someone else's queue.
What MCP does
The connection is called MCP, short for Model Context Protocol. It's a standard way to plug an AI model into the software a business already runs: the email platform, the analytics account, the website's code. Once a tool is connected, you describe the result you want and the model works through the steps in that tool, where you can see and check each one.
The second piece is what Claude Code calls a skill: a set of instructions and examples it reads before it starts. I trained one on our websites. It knows how our pages are built and how we write. When I ask for a new page, it drafts one in that style and sends it through GitHub to a staging branch. I review it there, and only then does it go live.
No developer, and no handover document. The review step is still mine, and it stays mine.
Why it matters to a small team
In a small marketing team the bottleneck is rarely ideas. It's the setup between the idea and the buyer: the list that needs splitting, the tracking that needs a tag, the landing page that has to exist before the campaign can start. When that work takes a sentence instead of a ticket, the idea reaches the market while it's still timely.
The question I can't put down
Every instruction I give, every correction I make and every draft I approve is a record of how our business thinks: who we sell to, what we say, and what we'd never say. I'm teaching these tools how we work, one task at a time.
So I've been asking my network two things. Are we training these models through the way we work? And how many of our software tools now pass our work through them? I suspect most teams couldn't answer the second question today.
If you're starting down this path, two habits help. Keep a written list of every tool you've connected to an AI model and what each connection is allowed to do. And read each tool's data settings yourself, account by account, rather than trusting a default you haven't seen.
The developer ticket I didn't write is easy to count. What the tools learned from watching me work is harder to count, and it's the number I'd most like to know.