The Setup
Your clients were drowning in ChatGPT outputs that sounded like every other ChatGPT output. Generic. Helpful, sure. But stripped of personality.
They’d ask the AI to draft a customer email and get something clinical. Ask it to write social copy and get the same tone as a bank’s LinkedIn post. The AI had no idea what their brand actually sounded like.
The problem wasn’t ChatGPT. ChatGPT doesn’t know who you are. It’s trained on the entire internet. You’re a blip.
The Build
The solution was custom GPTs — OpenAI’s way of letting you train an AI on your specific docs and instructions without needing to fine-tune a model yourself.
The process looked like this:
Step one: Voice audit. Everything your client had ever written. Old emails, marketing copy, social posts, internal docs, customer responses. I read it all. Not to memorize it, but to catch patterns. Sentence length. Favorite words. What they say when they’re annoyed versus excited. The gaps between what they think they sound like and how they actually sound.
Step two: Structured knowledge base. Clean, organized info the GPT could actually reference:
- Brand voice guide (tone, values, what not to say)
- Product catalog with descriptions and pricing
- FAQ section (real customer questions, how they answered them)
- Common objections and rebuttals
- Pricing justification
- Brand story and mission
Step three: Custom instructions. The GPT’s system prompt. Not just “be helpful” — something specific. “You’re the voice of [Company]. You’re warm but direct. You push back on bad ideas. You always mention the sustainability angle. You never promise features that don’t exist. When you’re not sure, say so.”
Then test. A lot. Ask it to draft customer responses, write a social post, explain a pricing decision. See where it hallucinates. Tighten the instructions.
The Mess
Custom GPTs aren’t magic. They’re really good first drafts. Sometimes they’re really bad first drafts.
The hallucinations were real. I’d give a GPT product info and it would confidently describe features that didn’t exist. You can reduce this with good instructions (“Do not invent features”), but you can’t eliminate it. The AI is always trying to be helpful, which sometimes means confabulating.
Knowledge cutoffs are their own nightmare. If your product launched after ChatGPT’s training data cut off, the GPT doesn’t know it exists. You have to explicitly feed it new info. And you have to keep feeding it. GPTs get stale.
Clients had expectations misalignment. They’d hand me a custom GPT and expect it to handle every customer question perfectly on day one. Some thought I was building Skynet instead of “a really good autocomplete for your brand.” I had to reset that repeatedly: this tool saves editing time. It doesn’t replace your voice — it speeds up your voice.
The other mess: OpenAI updates. When they change how custom GPTs work, they change what works for you. Maintenance isn’t optional. You can’t build a custom GPT, hand it off, and never think about it again.
The Result
The ones that worked were transformative.
A therapist could ask the GPT to draft a response to a potential client inquiry. Instead of staring at a blank screen for 20 minutes, the GPT delivered something in her voice — warm, boundaried, honest about her approach. She’d edit lightly and send it. Time saved: not hours, but enough. And the client got a response that actually sounded like the person they were considering hiring.
A product company used it to staff customer support before they could hire a person. The GPT answered common questions in their voice, on-brand, with accurate product details. Not perfect. But consistent. And when it wasn’t sure, it escalated to a human. The support team used it as a starting point and handled the judgment calls.
An agency used it to generate social copy first drafts. The AI knew their brand, their audience, their tone. It wasn’t writing the posts — their team was. But the blank page was gone.
The Takeaway
Custom AI isn’t about making AI perfect. It’s about making AI’s first draft close enough that editing beats writing from scratch.
If you’re spending 20 minutes staring at a blank email and then 10 minutes editing what an AI gave you, you won that trade. You moved faster. You got your brand into it. You didn’t rely on generic output.
The real insight: AI + your brand voice + your data = a tool that actually accelerates your work instead of just adding another tab to your browser.
What Changed Because of This
For clients, it meant relief. Less time on busywork copy. More time on strategy and actual thinking.
For me, it proved something I already believed: brand voice isn’t decorative. It’s practical. It’s the difference between a tool that saves time and a tool you ignore because it sounds like nobody.
It also taught me that custom AI needs a human in the loop. Every time. You’re not automating the human away — you’re automating the drag so the human can focus on the decisions that matter.
Want a custom AI that sounds like you? Book a Build Session — $350/90 minutes.