Business owners do not say no to AI. They say no to demos that do not feel like their business. Three prompts fix that: what it knows, how it sounds, and what it will not do. All three are below, ready to paste, along with the seven-minute demo order that turns "interesting" into "we need this."
You show a clean prototype. The owner watches, nods, and says "interesting" or "maybe next quarter." The meeting ends and nothing moves.
It was not the technology. It was the prompts. Most AI gets rules, and rules are generic. A business is not generic. It runs on exceptions: the client who pays late but always pays, the supplier you never call on a Monday, the phrase you never put in a contract because it once caused a six-month dispute.
Give the AI those exceptions, the owner's own voice, and a short list of lines it will not cross. Then demo it in that order. That is the whole method.
What Is the Triple Prompt?
The triple prompt is a three-part system prompt for business AI. Prompt one loads the exceptions the business runs on instead of the generic rules. Prompt two loads the owner's own vocabulary, learned from messages they wrote. Prompt three loads hard stops: the actions the AI never takes without asking first. Miss one and the whole thing falls apart. Together they produce an AI a stranger cannot replicate, because every input is proprietary to that one company.
Prompt One: What It Knows
Everybody gives AI rules. Nobody gives it the exceptions. The exceptions are where the business lives.
A rule says: "Always follow up with leads within 24 hours." The exception says: "Except partner referrals from Marco. Those go to me first, always." The rule is on every website. The exception is in the owner's head, and it is the only reason the follow-up works.
So do not start with the rules. Start with the things that go wrong.
Sit with the owner and ask one question: "Give me ten situations where the normal response would be wrong for your business." Then wait. The ones that come out fast are the ones that burned them before. Ten is the floor. A real business runs on a hundred of these.
Load every answer into the system prompt as an explicit exception, not as a rule. The difference matters. A rule tells the AI what to do by default. An exception tells it when the default is wrong.
Copy The Exceptions Prompt
Paste this at the top of the system prompt for the AI you are building, in Claude, Codex, or whichever runtime you use. Fill the brackets from the owner's answers.
Copy this.
You work inside [ACME PLUMBING], a [RESIDENTIAL PLUMBING COMPANY WITH 6 TECHNICIANS]. You have default rules. The exceptions below override them. Read the exceptions before every task. DEFAULT RULES - [FOLLOW UP WITH EVERY NEW LEAD WITHIN 24 HOURS.] - [QUOTE STANDARD JOBS FROM THE PRICE LIST.] - [SEND A PAYMENT REMINDER 7 DAYS AFTER THE DUE DATE.] EXCEPTIONS (each one beats the default it names) 1. [PARTNER REFERRALS FROM MARCO] go to [OWNER NAME] first, always. Do not contact them yourself. 2. [THE RIVERSIDE HOTEL] pays late but always pays. Never send them a reminder. Flag it to me instead. 3. Never call [SUPPLIER NAME] on a Monday. Queue it for Tuesday. 4. Never use the phrase ["FIXED PRICE"] in any quote or contract. 5. [ADD THE REST. TEN IS THE FLOOR.] HOW TO APPLY THEM - When a task matches an exception, apply the exception and say which number you applied, in one line. - When a task looks like it might match but you are not sure, stop and ask me before acting. - Never invent an exception. If it is not on this list, follow the default.
Your First Real Run
Take the exception the owner said first. Give the AI a task that trips it. If exception two is the hotel that pays late, hand the AI an overdue invoice from that hotel and tell it to run the reminders. The correct output is no reminder and a one-line flag: "Exception 2 applies. Riverside is on the always-pays list. Flagging to you."
If it sends the reminder anyway, the exception is written like a suggestion. Rewrite it as a hard line: "Never send them a reminder."
Done. The AI now knows one thing about this business that no template knows.
The Mistake That Makes It Fail
Loading the rules first. Write twenty rules and the owner reads them thinking any consultant could have written this, and they are right. Rules are public. Exceptions are proprietary. An AI that knows a hundred of them feels like it was built for this company, because it was. Start there.
Prompt Two: How It Sounds
Most AI conversations sound the same, because most AI learned the same generic voice. Now that anyone can mass-produce content with AI, the strongest weapon an owner has is sounding like themselves.
So teach it the voice from the source. Not a description of the tone. The owner's own messages.
Collect 50 messages the owner wrote. Emails, Slack messages, saved reply templates, the way they answer "how much does it cost." Then give the AI one rule: stick to this vocabulary.
| What to collect | Where to find it |
|---|---|
| Sent emails | Gmail or Outlook sent folder, last 90 days |
| Slack or Teams messages | Export from Settings > Workspace > Export data |
| Response templates | Any canned responses or saved drafts |
| Objection handling | Ask the owner: "How do you usually answer when someone says X?" |
50 is the threshold. Below 20, the voice is too thin and the AI drifts back to generic. Above 50, you get diminishing returns. Aim for 50.
The vocabulary rule has two lists. A short list of phrases the owner always uses, and a short list they never use. These are almost always business-specific: whether they say "customers," "clients," or "partners." Whether they name the price directly or ease into it. The greeting they always open with, or the one they would never write.
Copy The Voice Prompt
Copy this.
Below are [50] messages I wrote myself. Learn how I write from them: sentence length, how I open and close, how I refer to the people I work with, how I talk about money. RULES - Stick to this vocabulary. Do not upgrade my words to fancier ones. - Always use: ["CLIENTS", NEVER "CUSTOMERS"], ["LET'S GET THIS SORTED"], [FIRST NAME ONLY, NO "DEAR"] - Never use: ["GUARANTEE"], ["CIRCLE BACK"], ["I HOPE THIS EMAIL FINDS YOU WELL"] - Pricing tone: [DIRECT. THE NUMBER GOES IN THE FIRST TWO LINES.] - Greeting: ["HI [NAME]," EVERY TIME.] THE TEST Everything you write for me has to pass one test: someone on my team reads it next to something I wrote and cannot tell which of us wrote it. If you are not sure a line passes, rewrite it shorter. MY MESSAGES [PASTE THE 50 MESSAGES HERE, SEPARATED BY ---]
Your First Real Run
Ask the AI to write one reply the owner would normally write today, a quote follow-up or a scheduling change. Then run the team test. Send that reply and one the owner wrote to two people on their team. Ask: "Which one of us wrote this?"
If the team picks out the AI every time, the voice is off. Say: "Too formal. Look at messages 12 through 20 again and match the length." Keep refining until they cannot tell the difference. If the drafts still read like a machine after two rounds, run them through the humanizer pass before you touch the prompt again. That guide covers the tells. This one covers the voice.
Done. The AI now sounds like the person who signs the invoices.
The Mistake That Makes It Fail
Describing the voice instead of showing it. "Friendly but professional, concise, warm" produces the same AI voice every other company gets. Adjectives describe a voice. They do not reproduce it. Fifty real messages do.
Prompt Three: What It Won't Do
The most impressive thing your AI can do is stop when it should.
Every owner has heard the horror stories: the hallucinated contract, the email sent to the wrong person, the AI that kept charging a card after the customer cancelled. They are not afraid of AI being slow or expensive. They are afraid of AI doing something they cannot undo. And they are right to be.
So give the AI a list of everything it cannot do without asking first. Hard stops, not suggestions. There are three categories.
Money limits. Never commit more than X without approval. Never issue a refund over Y. Never send an invoice to a new client before the owner sees it.
Communication gates. Never send an outbound email to a list of more than 10 people without a review step. Never use the word "guarantee" in a response.
Platform boundaries. This AI handles WhatsApp. It does not touch the CRM. It does not post to social.
Write them as explicit rules. Not "try to avoid." Write "never do X without first asking the owner." Precision here is the product. A vague guardrail is not a guardrail.
Copy The Hard Stops Prompt
Copy this.
HARD STOPS These are not suggestions. Never do any of the following without first asking [OWNER NAME] and getting a yes in writing. MONEY LIMITS - Never commit more than [$500] on anything without approval. - Never issue a refund over [$100]. - Never send an invoice to a new client before [OWNER NAME] has seen it. COMMUNICATION GATES - Never send an outbound email to more than 10 people without a review step. - Never use the word "guarantee" in any message. - Never reply to [A LAWYER, A JOURNALIST, A REGULATOR]. Hand those to me untouched. PLATFORM BOUNDARIES - You handle [WHATSAPP AND EMAIL]. - You do not touch [THE CRM]. You do not post to [SOCIAL]. You do not log in to [THE BANK]. WHEN A REQUEST CROSSES A LINE Stop. Name the line you hit. Hand the draft back for review with this exact sentence: "I cannot send that without your approval. Here it is for review." Then wait.
Your First Real Run
Hand the AI a request that crosses a line on purpose. "Send this offer to the 40 leads from last month." The correct output is a stop, the name of the line, and the draft handed back for review. If it sends, the line is written as a preference. Rewrite it with "never" and "without first asking."
Done. The AI now refuses the one thing the owner was most afraid of.
The Mistake That Makes It Fail
Skipping this prompt because the demo already looks good. The owner is not thinking about how good it looks. They are thinking "what if it does something wrong." Until you answer that on screen, they will not say yes.
Once all three prompts pass their first real run, save them as one reusable skill so you never re-paste them. If you have never built one, this guide is the first build.
The Demo Sequence: Seven Minutes, Three Moves
The three prompts are the build. The demo order is what converts. Run it in this order and do not skip a step.
Minutes 0 to 2: context. Show the AI handling one of the owner's own exceptions. Put a real company name in the test data, a real product category, a real edge case. The line to open with: "This is not a chatbot. This is your Monday morning, already done." An unpaid-invoice chaser is a strong candidate for this slot. The invoice machine is the build for it.
Minutes 2 to 5: voice. Show the AI writing something in the owner's voice. Hand them the screen. Ask: "Does that sound like you?" If you ran the team test before the meeting, the answer is yes.
Minutes 5 to 7: guardrails. Show the AI declining a request that crosses a limit. "I cannot send that without your approval. Here it is for review." This is the moment the owner stops thinking about risk and starts thinking about implementation.
| Step | What it demonstrates | What the owner feels |
|---|---|---|
| Context | The AI knows their specific exceptions | "This is about us, not a demo" |
| Voice | The AI sounds like the owner | "This could actually represent us" |
| Guardrails | The AI stops at the right moment | "This is safe enough to deploy" |
Watch for the shift. In the first two minutes they are evaluating you. By minute seven the question changes from "what if it does something wrong" to "how soon can we implement this." That shift is the close.
A good demo is not impressive. It is a mirror. They walk in expecting a chatbot. You show them their own business, already handled. If you are still deciding what to sell in the first place, the installation business guide is what to sell. This one is how to demo it.
Honest: Where the Triple Prompt Does Not Save You
A prompt is a promise, not a lock. "Never commit more than $500" in a system prompt stops the AI most of the time, and it does nothing about a bug in the tool that moves the money. For money limits, back the prompt with a hard cap inside the tool itself, like a spend ceiling on the API key or an approval step in the payment flow. Show the owner both. They read the prompt. They trust the cap.
Below 20 messages, prompt two does not work. If the owner writes almost nothing themselves, the voice model has nothing to learn from, and no adjective fills that gap. Collect what exists, run the team test, and tell the owner the voice will sharpen as their sent folder grows.
If the owner will not give you ten exceptions, you do not have a client yet. Every real business has a hundred of them. An owner who cannot name ten has not decided to let you in. Do the interview anyway, but do not build the demo on generic data to fill the gap. A clean demo with fake data is exactly the demo that gets "interesting."
The triple prompt also does not make a weak AI product good. It makes a good one feel like it belongs to this company. If the underlying job (the follow-ups, the quotes, the reminders) is not worth automating, three prompts will not change the owner's answer.
Do This One Thing This Week
The order to build is the order to demo: what it knows, how it sounds, what it will not do. If you have time for only one, build prompt one. The exceptions are the part no competitor can copy, and the part the owner recognises in the first ten seconds.
Pick one owner you already know. Ask the one question: "Give me ten situations where the normal response would be wrong for your business." Write the answers into the exceptions prompt and run the first real run on the one they said first.
If it goes nowhere, you spent one conversation, and you are holding a list of ten exceptions that were never written down anywhere. That list alone was worth the meeting.