You’ve probably typed something into ChatGPT or Claude and felt a specific kind of disappointment. The words came out fine. They weren’t yours.
Nicole Noonan spent over a decade in experiential marketing, running activations for the NBA and Nike, chasing an Emmy in a production suite most people never see. Then, in 2021, she got what she calls the AI bug. She bought a paperback on Amazon about building custom GPTs and taught herself the rest. Marketing agencies told her AI was a fad and shut the door. She kept going anyway.
On this episode of A Good Pour: Summer of Good AI, Nicole and Kathryn talk about the real reason most AI output feels generic, and what it actually takes to train AI on your brand voice.
Why does AI feel broken for most people?
Nicole says close to ninety percent of people who feel AI doesn’t work for them are running into the same issue. The tool was never trained or seasoned. It doesn’t understand the business behind it, the workflows already in place, or the way that business actually talks.
That’s not a reason to give up on AI. It’s a reason to slow down before automating anything. Nicole’s rule is simple. Get the structure in place first, then let AI pick up the pieces that make sense: the scheduling, the follow-up, the routine outreach on LinkedIn nobody has two hours for. The strategic thinking stays with a person.
How does one business end up with more than one voice?
Kathryn and Nicole use the Nike comparison to make this concrete. There’s the broad company voice, the one behind “Just Do It.” Then there’s a separate voice for basketball, another for running, another for the athlete fronting a specific campaign.
Most small businesses have the same layers without realizing it. There’s the company voice. There’s the founder’s voice, which usually needs to sound more like a thought leader than a brand account. If AI only ever learns one of those voices, it starts flattening the other one out. The fix isn’t a longer prompt. It’s training each voice on its own, on purpose.
What does an MCP actually do?
Nicole’s plainest explanation: an MCP works like the phone system businesses used to run everything through. One hub, all the connections routed through it. That’s what an MCP does for AI. It routes requests to the right tool, whether that’s Gmail, a project management system, or a reporting dashboard, and it learns which one to pull based on how it’s been trained.
Most major platforms already have some version of this built in, whether that’s Claude’s native integrations or dedicated MCP tools like Zapier or N8N. The barrier to entry is lower than it looks. The bigger question is what you’re connecting, and why.
Where’s the real AI risk hiding?
Security, not automation, is the piece Nicole sees people skip. She compares it to a gated community. It feels safe because the settings say private. But someone determined to get in usually finds a way, and most people never stop to think about what they’re feeding into a public AI tool in the first place: private reports, internal methods, client information.
The same logic applies to which tools a team is allowed to use. If a business wouldn’t let an employee choose their own email platform, it shouldn’t let them choose an unapproved AI tool either, especially on a personal account nobody’s trained.
Use This Today
Pick one task you already do by hand every week and ask three questions before you touch an automation. How many times do I do this? What does it cost me in time? Would I actually miss it if it disappeared? Nicole runs this exact check with clients before automating anything. If the answer doesn’t justify the setup, leave it alone for now.
Not every task needs a system built around it. Some of what feels behind is actually just the noise of every tool promising to fix everything at once. Pick one thing. Build on it. That’s enough for this week.
Next step
If you want a starting point for training your team, GCM’s free AI Policy Document walks through where to draw the lines before AI touches your business voice or your data. [LINK: AI Policy Document]
FAQ
Why doesn’t AI sound like me?
Most AI hasn’t been trained on your specific voice, only given a single prompt and expected to guess. Feeding it real examples of your writing, your word choices, and the difference between your company voice and your personal voice fixes most of the gap.
What is an MCP in AI?
An MCP works like a phone system for AI tools. It routes requests to the right connected app, whether that’s email, a calendar, or a reporting tool, based on how it’s been trained to work.
How do I know if I should automate a task?
Start with how often you do it, how much time it costs, and whether the result actually matters if it’s missed. If a task doesn’t clear that bar, it’s not worth automating yet.
Is it safe to put business information into ChatGPT or Claude?
Not without thinking it through first. Private reports, client information, and internal processes can end up stored or referenced in ways a business didn’t intend, so treat AI tools with the same access rules as any other software a team uses.
Training AI on your voice isn’t a technical project. It’s the same work Nicole did back when she got the AI bug in 2021: figure out what makes the business sound like itself, then teach the tool to keep up.