TL;DR: Set up a Zapier flow that pulls new Gmail messages, sends them to ChatGPT, and saves a draft reply back to the thread. Never auto-send. Use a system prompt with four clear decision rules. Do that and you'll cut inbox time by 70% without losing your voice.
If you run a bar, a restaurant, or any service business, email is one of those things that doesn't look like a real problem until it is. You're putting out fires until 2pm, you finally sit down, and there are 60 unread messages. A vendor following up. A guest complaint. A booking request. A distributor rep who sent the same thing twice. A team member forwarding something that probably didn't need to be forwarded.
Most operators deal with this by triaging in their heads and answering the urgent ones, letting the rest pile up, and feeling vaguely guilty about it on a rolling basis. The backlog becomes ambient stress. You're not ignoring people. You're just drowning.
The AI email tools that most people try first make it worse, not better. They sound corporate. They use phrases like "Thank you for reaching out" and "We apologize for any inconvenience." Nobody who works in hospitality talks like that. Your guests, your vendors, your partners, they can tell immediately that a robot wrote it. And now you've got a trust problem on top of a time problem.
What I built is different. My assistant, Jessi, drafts replies. She never sends. I open the thread, skim the draft, edit the 20% that needs it, hit send. That's it. The blank page problem is gone. I'm still in the loop on every message. The voice sounds like my business, not like a corporate FAQ.
Auto-send is where this goes wrong. Everyone who tries to fully automate their inbox eventually has a horror story. A guest complaint that got a chirpy templated response. A vendor email that got answered with context that was completely off. A situation that needed judgment and got a formula instead.
The draft model fixes that. The AI does the hard part, which is staring at a blank reply box with someone waiting. You do the easy part, which is reading a draft that's already 80% right and tweaking the last 20%. Per email, that's about 10 seconds of real work versus 3 minutes.
Over 60 emails, that math is significant.
The other thing the draft model does is keep you accountable. You're still reading every message. You still catch the weird ones, the urgent ones, the ones that need a real conversation instead of a reply. The system handles the volume. You handle the exceptions. That's how it's supposed to work.
The core setup is three steps in Zapier.
Step one: Trigger. Use "New Email Matching Search" in the Gmail app. Set the search criteria to label:inbox. This catches every new message that lands in your inbox.
Step two: Action. Use ChatGPT, "Send Message." This is where your system prompt lives. Paste your four rules here (more on those below). In the message field, pass through the sender name, the sender email, and the full email body from the trigger step. Give the model enough context to actually write something useful.
Step three: Action. Use Gmail, "Create Draft Reply." Map the draft body to ChatGPT's output. Map the thread ID from the original email trigger. The draft lands in the right thread, attributed to the right conversation. You open Gmail, go to your drafts, and you've got replies waiting for you.
That's the whole thing. Three steps. Takes about 20 minutes to set up the first time.
Without this, the assistant will try to draft a reply to your own emails. You reply to yourself on a thread, it tries to respond. Someone on your team emails you, it drafts an external-sounding reply. That's messy.
Add a filter between the Gmail trigger and the ChatGPT step. The condition: "From Email" does not contain @yourdomain.com. If the sender's email includes your own domain, the zap stops. If it's from anyone outside your organization, it continues.
One filter. Massive difference in how the system behaves.
This is what most people skip. They drop a one-line instruction into the system prompt and wonder why the drafts sound generic. The system prompt is where the personality lives. Take the time to write it properly.
Mine has four rules, and they're specific enough that Jessi actually applies them differently depending on the situation.
Respond with urgency. Kindness is the default tone, not something that gets turned on for upset guests. Every email, regardless of content, gets warmth built in. The draft shouldn't feel like it was written by someone who's annoyed that they have to respond.
Always provide a solution. Never use language that buys time without offering anything. "I'll look into this and get back to you" is a stall. "Here's what I can do right now, and here's what I'll confirm for you by tomorrow" is a solution. The instructions in the prompt are explicit: no stall language, ever.
This is the rule nobody else talks about. If someone sends you 30 characters in all caps, they're frustrated. A 30-character reply matches the form but misses the situation. You go longer to de-escalate, acknowledge what's happening, and give them something real. If someone sends a friendly 200-word message, you don't reply with three sentences. You match their energy.
The "roughly match length" guideline I use is within 100 characters of the sender's message length. That's not rigid, it's a calibration. It keeps the drafts from feeling curt when someone put effort in, or rambling when someone just asked a quick question.
The draft closes with Jessi's name, not mine. This matters. I'll explain why in the next section.
Transparency. That's the whole reason.
When someone gets a reply from Jessi, they know it's an assistant. I'm not pretending to have personally written every message. That honesty actually increases trust, not decreases it. People appreciate knowing their message was handled promptly by a capable assistant who operates within clear guidelines. That's not a downgrade from "Jason wrote this." That's professional.
The alternative, where the AI drafts in your name and you pass it off as your own writing, is where things get uncomfortable. You're asking the model to impersonate you. The guest asks a follow-up and now you have to remember what "you" said. The vendor references the conversation and you're scrambling to find the thread. It's a paper-thin deception that's more work to maintain than it's worth.
Name the assistant. Let the assistant be the assistant. Your credibility comes from the quality of the operation, not from pretending you typed every email yourself.
The setup runs. The drafts show up. Some of them are great. Some aren't. Here's what to look for and what to do about it.
Tone drift. The drafts start sounding progressively more formal or more casual than you want. Fix it by adding a line to the system prompt: "Match the tone of [your business name]. Warm, direct, not corporate." Specificity helps more than generality.
Over-apologizing. Common with the empathy rule if you don't constrain it. Drafts that open with "I'm so sorry for any trouble this may have caused" for a routine booking request. Add to the prompt: "Apologize only when there's a clear mistake or failure. Do not pre-emptively apologize."
Wrong length calibration. If the drafts are consistently too long or too short, adjust the prompt with explicit guidance: "Target 150-250 words for most replies unless the sender's message is significantly shorter."
Sharing information the assistant shouldn't have. Jessi doesn't know your calendar, your current inventory, your pricing exceptions, or your private operational details. If the prompt has access to sensitive context, it can leak into drafts. Keep the system prompt focused on response style, not internal data.
Scheduling without checking availability. This is the big one. The assistant will confidently suggest times it knows nothing about. Add a hard rule: "Never confirm specific dates, times, or availability. Always direct scheduling requests to [your booking link or preferred method]."
Before I built this, email was 60 to 90 minutes of my day, fragmented across whenever I could get to it. Now it's about 20 minutes, usually in one sitting in the morning.
The draft is right about 80% of the time. Maybe more. The other 20% is a sentence that's slightly off, a detail that needs updating, or a judgment call that I actually want to make myself. Editing is fast. The blank page was the slow part.
Over five days, that's 5 or more hours back. Hours I was spending on email that nobody was paying me for, that wasn't growing the business, that was pure administrative drag.
That time goes somewhere better now.
Does the AI actually send the emails?
No. The system creates a draft in Gmail. It lands in your Drafts folder, attached to the right email thread. You review it, edit it if needed, and hit send yourself. Nothing goes out without your approval. That's not a limitation. That's the design.
What if someone replies to the assistant asking if they're a real person?
They can ask Jessi directly. Jessi's instruction is to be honest: she's an AI assistant. That's it. No elaborate explanation, no apology. "Yes, I'm an AI assistant. Jason reviews and sends all replies." It's a non-issue in practice. Most people don't ask.
Can I use Claude or a different model instead of ChatGPT?
Yes. Zapier has integrations for several AI models. Claude handles nuance well and can be more natural in conversational tone. The system prompt logic, the four rules, the structure, all of it transfers directly. The model is interchangeable. The prompt is what matters.
What if the draft is wrong or off-tone?
Edit it. That's the whole point of the draft model. Nothing is locked in. When you see patterns in the errors, add a line to the system prompt that addresses them directly. Most tone issues get solved with one specific instruction. The system learns the correction for every future draft.
How do I stop it from replying to promotional email?
Two options. Refine the Gmail search criteria in the trigger step. Instead of just label:inbox, add -label:promotions or -category:promotions to exclude Gmail's promotional tab. Alternatively, add a filter step that checks the email category and stops the zap for anything flagged as marketing or automated. Most operators find the category filter is enough.
If you run a service business, consulting firm, or multi-location operation and want to see what a fully built AI back office looks like, the KMS demo is at kmsops.com or you can book directly at kmsops.com/connect.
If you're a restaurant or bar operator and want to see how this fits into a broader operations system, that's ASM Command.
More playbooks like this one come out in The Ops Wire newsletter. You can subscribe at theopswire.substack.com.
Jason Littrell spent 10 years behind the bar in NYC, including Death & Co, and served as USBG NYC president. He now runs his hospitality consulting firm entirely on AI. He hosts the Hospitality Strategy Lab podcast and writes The Ops Wire newsletter.
Jason