Last July, I sat down with Michelle, a personal chef based in Austin who cooks for over 10 clients a week.
She’s good at what she does. Custom menus, dietary restrictions tracked, clients with strong preferences about spice levels and proteins, and what they ate last week. Real relationship-based work.
And every single week, she was spending four hours writing menus by hand.
I asked her what that process looked like. She’d open her client list, pull up each person’s notes, think through what she’d made for them recently, factor in what was seasonal or on sale, and then type out a custom menu for each client from scratch.
Every week. Same process. Four hours gone.
When she said that, I didn’t immediately say, “Here’s the AI tool for that.” I asked more questions first. What format do the menus go out in? Where do you keep the client preferences? How much does the menu vary week to week?
That diagnostic step matters more than most people think. It’s the difference between building something that works and building something that looks good in a demo but fails on Tuesday.
The First Attempt (And Why It Broke)
Michelle had already tried to fix this herself before we talked.
She built a ChatGPT prompt that was supposed to do everything: read the client preferences, remember past menus, generate new ones, and export them to CSV all at once.
It kept failing. Forgetting which client was which. Messing up the CSV. Generating the wrong proteins for someone with dietary restrictions.
She figured she’d done something wrong with the prompt. But the prompt was fine. The architecture was the problem.
One agent trying to do four separate jobs at once is like asking someone to be the chef, waiter, cashier, and dishwasher simultaneously. They’ll do all of it badly. There’s a design principle I use with every client now: one agent, one job.
When you try to cram too much into a single AI prompt, you’re overloading the context window, creating conflicting instructions, and making it nearly impossible for the model to know which task should take priority when they conflict. The result is exactly what Michelle experienced — inconsistent, unreliable output that you can’t trust.
What We Actually Built
We spent about 90 minutes mapping the workflow together before touching any AI tool.
First, we figured out the data. She already had a spreadsheet with client preference tabs — one tab per client, with notes on their allergies, favorite proteins, what they’d eaten recently, and any restrictions. That spreadsheet was the source of truth we needed.
Then we designed backwards from the output. What does a finished menu actually look like? How long is it? What format does it need to be in? Once we knew exactly what “done” looked like, building toward it got a lot simpler.
The workflow ended up with two focused steps:
- Read the client preference tab and pull the relevant data
- Generate a personalized weekly menu based on those preferences
That’s it. No CSV export in the same step. No memory of past menus baked into the same prompt. Just two clean, focused tasks.
Michelle reviews the output, makes any adjustments that feel off (she still knows her clients better than any AI does), and sends them out.
The whole thing now takes about 30 minutes instead of four hours.
The Dual-Brand Problem
Michelle also runs two different lines of business under her name — a luxury catering arm and the personal chef meal service. They have completely different voices. The catering side is formal, polished, upscale. The personal chef side is warmer, more casual, like a friend who happens to cook for you.
She’d been writing all her client communications by hand to manage that tonal difference. Which made sense — she just didn’t have another option.
So we added a second piece to the workflow: she used ChatGPT to analyze her existing content for each line of business and write a brand voice prompt that captured the tone. Then we imported those prompts into her AI agent’s system settings.
Now when the agent generates a menu or a client update for the catering side, it automatically writes in the right voice. Same for the personal chef side. Two brands. Two distinct voices. Zero extra effort on her end.
This “use AI to write the prompt for another AI” approach sounds a bit meta, but it works really well for service businesses with multiple client segments or brand voices.
What Made This Actually Work
A few things made this project go smoothly that I want to call out because they’re not obvious.
She had clean data. The client preference spreadsheet was organized and up to date. If that hadn’t existed, we would have spent the first hour just figuring out where her client information lived. Data centralization isn’t glamorous, but it’s what makes automation possible.
We didn’t start with the tool. We started with the workflow. What’s the desired output? What’s the trigger? What are the steps a human would take? I call this designing backwards — start at the end state and work backward. It keeps you from building a solution in search of a problem.
She stayed in the loop. Michelle still reviews every menu before it goes out. The AI isn’t making final calls — it’s doing the drafting work. That human review step means clients still get her judgment, and she catches anything the AI gets wrong before it matters.
That last point is worth sitting with. The goal wasn’t to remove her from the process. It was to remove the part of the process that was just mechanical repetition, so she could focus on the part that actually requires her expertise.
Your Version of This
Almost every service business has a version of Michelle’s four-hour menu writing task. Some repeating process that requires your knowledge, but not your active presence every single time.
It might be client intake forms. Proposal generation. Weekly check-in emails. Social media captions. Progress reports.
The pattern is almost always the same: you know exactly what needs to go in, you’ve done it dozens of times, and yet you’re still sitting there typing it from scratch every week.
The fix usually isn’t complicated. It’s mostly about designing the workflow right before touching any AI tool.
If you’re curious what this might look like for your business, I run AI workshops where we actually build these systems together — not just talk about them. Check out my upcoming workshops at asianefficiency.com and see if there’s one that fits.
Or just try mapping your own version of “the four-hour task” this week. Write down the steps a human would take. Figure out where the data lives. Then ask yourself: which of these steps actually needs me?
That question alone might get you most of the way there.
