TL;DR: You can train a regular AI assistant to write in your voice using materials you already have. It takes about 30 minutes of prep, two text files, and a willingness to test outputs until they stop making you cringe. The result is a drafting tool that sounds like you instead of a press release.
Most hospitality operators I know hate writing. Not thinking, not communicating, not connecting with guests or staff. Writing. The blank page, the Instagram caption that needs to go up before service, the response to a passive-aggressive Yelp review at 11pm when you are three covers deep and running one server short.
Generic AI does not solve it, because generic AI sounds like it was written by someone who has never held a shift drink or had to talk a line cook off a ledge at 6:45 on a Friday.
There is a fix. People in the marketing world have been calling it an "AI clone." The term sounds creepy. It is not. Moving on.
What it actually is: a version of an AI assistant that has been given enough context about how you talk, what you care about, and what you refuse to say, that it produces drafts in your voice instead of generic corporate-speak. You still review everything before it goes anywhere. The clone drafts. You ship.
Not a deepfake. Not a bot pretending to be you in DMs. Not something that replies to guests without your eyes on it first.
An AI clone, for this purpose, is two things: a detailed voice profile and a task prompt. Together they give a standard AI tool (ChatGPT, Claude, whatever you use) enough context to write the way you talk. It is not magic. It is specificity.
Marketing agencies have been doing this for clients for a while now. Their version usually starts with 20 to 30 minutes of polished video. Record yourself, upload the files, analyze the patterns, build the prompt. That works if you have been creating content for years and have a library to pull from.
Most hospitality operators do not have that. You have voice memos from the parking lot after a bad Saturday, a folder of staff texts explaining the new wine program, and maybe one podcast appearance you have not listened to since. That is enough.
The clone handles writing tasks that follow a pattern: review responses, staff updates, vendor follow-ups, social captions, newsletter blurbs. It does not handle anything that requires reading the room in real time or being a person. We will come back to those limits.
Every tutorial I have seen on this topic starts with video. Record yourself presenting, upload the files, let the AI study your delivery. That process works if you have been producing content consistently.
Most operators have not. Beverage directors are not running YouTube channels. Bar owners are not doing weekly Instagram Lives. GMs are not hosting podcasts.
The goal is not video. The goal is enough raw examples of how you actually communicate to teach an AI the patterns. That material is everywhere in your operation. Voice memos count. Staff texts count. Old emails count. One podcast appearance transcript, if you have it, is worth ten polished captions because you are talking off the cuff and the transcript shows how you actually build an argument.
Thirty to sixty minutes of this is enough. You do not need a library. You need a representative sample.
Here is what to pull together before you build anything:
3 to 5 voice memos talking through real situations. A bad comp call. A staff conflict you resolved. A menu change that worked. A night that went sideways and why. Do not script these. Just talk. The messiness is the point.
Your last 10 to 20 written messages to staff. Group texts, Slack messages, email updates, training notes. Anything you wrote when you were trying to explain something clearly to people who know you.
1 to 2 transcripts from any podcast or panel appearance. Even a 20-minute interview is plenty. Most hosts will send a transcript if you ask. If not, paste the auto-captions from YouTube and clean up the obvious errors.
One short bio paragraph you have actually written and approved. Not a LinkedIn summary someone else drafted for you. The version you wrote yourself, in your own words, even if it is rough.
One list of words and phrases you never use. This is the file most people skip. It is also the most important one you will build. Every operator has phrases they find condescending, words that feel wrong, ways of talking about the business that make them wince. Write them down. The clone needs to know what you refuse to sound like as much as what you do sound like.
Once you have your raw material, you build two things.
The voice profile prompt. This is the "who I am" file. It tells the AI how you communicate, who you are talking to, and what you avoid. A working voice profile covers: your background and why it matters to your audience, the tone you write in (direct, warm, peer-to-peer, not formal), the words and phrases you never use, and two or three examples of writing you are proud of.
In plain language it looks like: "You write like a working operator talking to another operator. You do not explain basics. You do not hedge. You would never write 'per my last email.' You would never say something is 'transformative.'"
Vague instructions produce vague output.
The task prompt. Every time you want a draft, you pair your voice profile with a short, specific task prompt: what you need, who it is for, any details to include.
For a Yelp response it might be: "Write a 75-word response to a 3-star review. Guest mentioned the wait but said the food was good. Acknowledge the wait, apologize without overdoing it, and invite them back for a specific reason."
Voice profile stays constant. Task prompt changes every time.
Before you send anything a clone drafted, run it through five jobs you already know the answer to.
A staff thank-you note after a hard week. A Yelp reply to a 3-star review where the food was great but the service was slow. A vendor follow-up after a late delivery. A 60-word Instagram caption about a new menu item you are proud of. A 200-word email pitching yourself for a podcast or panel appearance.
You know what your answers to these should feel like. Read the drafts and notice where they go wrong.
If the output makes you cringe, your source material is too thin. Go back and add more voice memos, more staff texts, more real writing. The AI is pattern-matching on what you gave it. If the patterns are sparse, the output will be generic.
If the output sounds like a LinkedIn motivational poster ("This team embodies the spirit of hospitality in everything they do"), your "never use" list is too short. Add the specific phrases that showed up in the draft that you would never actually write. Feed the corrected version back as an example of what not to do.
This is iterative. The first draft is not the final one. You are training it the same way you train a new hire: give feedback, show examples of what good looks like, correct the mistakes before they become habits.
A writing assistant that sounds like you is not a replacement for you making decisions.
Do not let the clone reply to angry guests directly. It cannot read tone, history, or context. A human moment handled badly by an AI proxy is worse than a slow human response.
Do not let it negotiate prices. Those conversations require judgment calls it is not equipped to make.
Do not let it sign anything, or draft anything that implies a commitment without your full attention on it.
Do not let it write real apologies. Not the "sorry for the inconvenience" variety. The kind that matters, where something went genuinely wrong and someone deserves to hear from a person. People notice the difference.
The clone handles volume. You handle stakes.
There is one test that matters more than reading the output yourself.
Send a clone-drafted message to a staff member or industry peer who knows how you communicate well enough to notice when something is off. Do not tell them it was drafted by AI. Just send it.
If they hesitate, respond differently, or ask if you are doing okay, keep training. They noticed even if they cannot name it.
If they respond the way they always respond to you, the clone is working. They did not notice because there was nothing to notice.
That is the finish line.
How long does it take to build one of these?
Raw material collection takes about 30 minutes if you have voice memos and old texts to pull from. Writing the voice profile prompt takes another 30 to 45 minutes. Stress-testing and refinement adds an hour or two over the first week. Call it three to four hours total before you have something you trust.
Which AI tool should I use?
The two most common are ChatGPT and Claude. Both work. The differences matter more at the edges than for everyday writing like review responses and staff communications. Start with whichever one you already use.
Won't this make my writing feel fake?
Only if the source material is fake. Build the voice profile from real writing and real recordings of how you actually talk, and the output will feel like you because it is modeled on you. The fakeness in most AI writing comes from prompting without context. You are providing context.
What if my staff finds out I use AI to draft messages?
Most will not care, and some will ask how to do it. The conversations that matter are in person. A well-drafted written update is not less real because you had help with the phrasing. You still read it, approved it, and sent it.
Can I sell access to my clone?
For hospitality, no. Your clients and guests are paying for your judgment, your track record, your presence. A writing assistant built on your voice is a production tool for your own business, not a product.
Three steps. Do them this week or do not bother.
Step 1: Record 3 voice memos. Real situations, not scripted. Walk to your car after a shift and talk through what happened. Talk through a staff issue you are sitting with. Talk through something that worked and why. Do not edit. Do not start over. Just talk.
Step 2: Write your "never use" list before you write anything else. Open a notes app right now and write down ten phrases that would embarrass you to have your name on. Industry clichés, corporate-speak, words that feel wrong for how you actually talk. This list will save you more time than anything else you build.
Step 3: If you want help building this into your operation, see ASM Command at https://jlittrell.com or KMS at https://kmsops.com
More playbooks like this go out in The Ops Wire newsletter at https://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