AI Assistant
How to set up an AI assistant: a practical starting guide
The plumber spends thirty minutes on a Thursday afternoon setting up an AI tool to handle his inquiry emails. He gives it a generic description of his services, turns it on, and goes back to work. The next morning, it has sent three replies with wrong service areas and a pricing range that is off by forty percent.
The setup is where most AI assistants fail. Not the AI. The setup. This article covers what a working setup actually looks like: what to build before you turn anything on, how to establish the gate that keeps the assistant from acting without your approval, and where the common failures happen.
Before you set up anything: build the context
An AI assistant works from the context you give it. It knows nothing about your business by default. That means the first step is not configuring a tool. The first step is writing down what the tool needs to know.
Call this your business context document. It does not need to be long or formal. It needs to be accurate and specific. Here is what it should cover:
What you do. Not a marketing paragraph. A factual description. Service types, how they work, how long they take, what they cost. "Residential HVAC service and installation in the greater Columbus area. Service calls start at $125. New system installation starts at $4,500 depending on size and system type." That is the level of specificity that produces accurate answers.
Your service area. The exact geography. Cities, zip codes, or counties you do and don't serve. An AI assistant that doesn't know your service area will book appointments in places you don't go.
Your schedule and availability. When you are available for appointments. How far in advance to book. How long different job types take. Blocked times. If this changes week to week, you need a way to update it.
Your standing rules. How you handle cancellations. Your payment terms. Whether you offer financing. What warrants a callback versus a message. The things you do consistently that you would tell a new employee on their first day.
How you want to sound. The tone your messages should have. Formal or casual. How you address customers. Whether you use first names. A few examples of messages you have sent that hit the right register.
Writing this down takes two to three hours if you do it seriously. It is the most important work in the setup. Everything the assistant produces comes from this foundation.
The gate: what the assistant cannot do without your yes
The second most important part of the setup is the approval gate. Define clearly what the assistant can do on its own and what requires your review before it happens.
A useful default structure for most small businesses:
Draft and hold. The assistant drafts a reply, queues it for your review, and sends only after you approve. This applies to: anything that goes to a customer for the first time, anything involving pricing, anything involving a commitment you haven't made yet, anything that doesn't fit the normal pattern.
Draft and send. The assistant sends directly without your review. This applies to: routine acknowledgments ("We got your message, we'll be in touch shortly"), post-job thank-you messages using a template you've approved, scheduled follow-ups at fixed intervals you've set up.
Start with draft-and-hold for everything. Move items to draft-and-send only after you have seen enough output to trust them. The goal is to build trust with the system over time, not to trust it out of the box.
Whatever gate structure you set up, make sure the AI cannot bypass it by being prompted to. The gate should be a rule in the system, not a request to the AI to be careful. AI that can be prompted around a gate will eventually be.
Start with one task type
Pick the single most repetitive writing task in your business. New inquiry responses. Post-job follow-up messages. Scheduling confirmations. One task type.
Set up the assistant for that task and only that task. Run it for two weeks. Read every output. Correct what's wrong and update your context document based on what the corrections reveal. You will discover three or four things you forgot to include. Fix them.
After two weeks, you have a working assistant for one task that you understand and trust. Add the next task. Repeat.
The businesses that succeed with AI assistants build them one workflow at a time. The ones that fail try to hand over everything at once and abandon the whole thing when the first thing goes wrong.
Where setups fail
Vague context. "We provide HVAC services in Ohio" produces much worse output than the specific version. Vague input, vague output. The assistant will fill gaps with guesses and the guesses will be wrong.
No gate. Any AI assistant that sends things, books things, or commits to things without an approval step is a liability. It will do something you did not intend, at an inconvenient moment, to a customer you needed to handle carefully. The gate is not optional.
Not reading the output. Every output the assistant produces for the first month should be read before it goes out. Even if it usually looks right. The one that is wrong will be the one you didn't read.
Static context. Your business changes. Prices change. Availability changes. Policies change. If your context document does not get updated when those things change, the assistant will keep answering based on what was true six months ago. Build a habit of updating it when anything material changes.
Wrong tool for the task. General AI tools (Claude, ChatGPT) are excellent for drafting work where you paste in the context and the prompt each time. They are not set-it-and-forget-it systems. An AI assistant product that maintains persistent context and a workflow is a different category of tool. Know which one you are using and what it actually does.
What it looks like when it works
When the setup is solid, a working AI assistant looks like this: a new inquiry comes in at 8 PM. The assistant drafts a response using your real service information and your real tone. It queues the draft for your review. You check it in the morning, approve it or adjust one line, and it sends. The customer heard from you the same evening they inquired, and the reply is accurate.
The post-job follow-up goes out two days after the job. The invoice draft shows up with the correct line items from the job notes you entered. The scheduling message goes out with the right address and time.
None of this requires you to be at a keyboard doing it. All of it required you to build the context and set up the gate.
The AI assistant that works is not the one with the most features. It is the one with the most accurate context and the tightest gate. Start there. The rest follows.
Build it right the first time.
The AI Field Guides for business owners and solopreneurs cover the full setup: what to write in your context document, which workflows to start with, and how to build an assistant that holds up under daily use.
Guides for business owners → See all guidesGo deeper
AI Field Guides from $9.99
Practical field manuals for real operators. Each guide covers one tool or workflow in full: the setup, the prompts, the limits, and the results. Available on Amazon and Kindle Unlimited.