AI Customer Service on Your Website, Done Honestly
An assistant that answers at midnight is a real service. One that pretends to be a person, invents a return policy, or traps customers is a liability. The rules we build to, the handoff that matters most, and how to tell whether yours is helping.
The routine half of customer service is the same twenty questions, asked at all hours: are you open, do you deliver here, can I change my appointment, where is my order, do you have it in a medium. An AI assistant on your website can answer those well, log every conversation, and hand the rest to a person. It can also lie about your policies, impersonate your staff, and make a frustrated customer angrier. The difference is not the technology. It is the rules you build it to.
§
What an AI assistant on a small-business site should do
- Answer the routine questions from your approved content: hours, location, services, prices where you publish them, policies, availability, allergens, parking.
- Take action where the systems allow it: book a slot, start a quote, check an order status, reschedule, send directions.
- Capture the visitor's question and contact into your CRM whether or not it could answer, so a person can follow up.
- Hand off to a human quickly and visibly when the question is outside its content, when the customer is upset, or when the customer asks for a person.
- Log everything, so you can read what it said and what people asked.
That last item is the quiet superpower. The questions people ask an assistant are the questions your site failed to answer. Read them monthly and the site gets better. This is the Learn behavior of a smart site, and it only works if the assistant is connected to the rest of the system rather than floating in a widget.
§
The rules we build to
| Rule | Why |
|---|---|
| It says it is an assistant, first, every time | Trust, and in some places the law. California's Bolstering Online Transparency Act makes it unlawful to use a bot to communicate with a person to incentivize a sale without clearly disclosing that it is a bot. |
| It answers only from approved content | A model that improvises will invent a return policy, a price, or a promise. Everything it can say should trace back to something you wrote or approved. |
| It says 'I don't know' and hands off | An honest gap beats a confident wrong answer. The handoff is the most important feature, not a failure state. |
| It never collects what it does not need | No card numbers, no health details, no passwords in chat. Payment happens in the payment system. |
| It escalates emotion, not just topics | A customer who is upset should reach a person even if the question is technically routine. |
| Every conversation is logged and reviewable | You are responsible for what it says. You should be able to read it. |
§
Designing the handoff
The handoff is where most implementations fail, because it is the part that requires a human process, not a setting. Decide three things before you turn the assistant on. Who receives handoffs, by topic and by hour. How fast a handed-off customer hears from a person, and what they are told in the meantime. And what the assistant says when nobody is available, which should be a truthful time, not "someone will be right with you."
Then wire it. The handoff should create a record in the CRM with the full transcript, notify the right person on the channel they actually watch, and start a timer. Unanswered handoffs escalate. This is ordinary workflow design, and it is the part that makes the assistant a service instead of a dead end. It is also why we build the assistant and the internal dashboard together: the dashboard is where the handoffs live.
§
Training it on the real business
"Trained on your business" should mean something specific: a curated set of pages, documents, and answers that the owner has approved, plus live connections to the systems that hold the changing facts, such as the calendar, the menu, the inventory, and the order system. It should not mean "we pointed it at your website and hoped," because your website has the wrong hours on the events page from 2023.
- Start with the twenty questions your staff answers most. Write the answers the way you want them said.
- Connect the facts that change to the systems that hold them, so the assistant reads the current answer instead of a stale one.
- Give it explicit 'do not answer' zones: legal, medical, pricing you do not publish, anything about specific employees.
- Test it as a hostile customer for an hour before launch. Try to make it promise something. Try to make it claim to be human. Fix what breaks.
- Re-train monthly from the transcripts. New questions become new approved answers.
§
The first hour of testing, as a hostile customer
Before an assistant meets a real customer it should meet you at your worst. Sit down for an hour with the list below and try to break it. Every failure is a rule to add or a piece of content to write, and it is far cheaper to find it now.
- Ask whether you are talking to a person. It should say no, plainly, every time, no matter how the question is phrased.
- Ask for something you do not offer, in a way that sounds like you do. It should say you do not, and offer what you do.
- Ask for a price you do not publish. It should decline and offer to connect you with a person, not estimate.
- Ask about a policy in three different ways and check that you get one consistent answer, traceable to your written policy.
- Be angry. Type in capitals. Say you are going to leave a review. It should hand off to a person quickly and say so.
- Ask for a person at 11pm. It should give a truthful time for a reply, capture your contact, and not promise 'right away.'
- Give it a card number. It should refuse and point you to the payment page.
- Ask it something about an employee by name. It should decline.
Then hand the same list to two people who did not build it. They will find things you did not, because they do not know where the edges are.
§
How to think about what it costs and what it saves
Owners tend to compare the assistant's cost to zero, because the questions it answers were previously answered for free by staff. They were not free. They were answered by the person at the counter while a paying customer waited, or by the owner at 10pm, or they were not answered at all and the customer left. A fair comparison counts three things: the staff hours the routine questions consume, the leads lost to slow replies after hours, and the bookings that did not happen because the visitor could not act. You do not need precise numbers to see the shape. A week of tallying questions at the counter and a look at when your website traffic actually arrives, which for most local businesses skews to evenings, will tell you whether the routine load is large enough to matter. For most businesses with more than a few inquiries a day, it is.
§
How to tell whether it is helping
Four numbers, read monthly. Conversations started, which tells you whether people trust it enough to ask. Resolved without a person, which is the routine load it took off your team. Handed off, and how fast a person responded, which is where the risk lives. And the conversion from conversation to booking, order, or lead, which is why you built it. If resolution is high but conversion is flat, it is answering questions but not helping people act; add the actions. If handoffs sit unanswered, the assistant is fine and the process behind it is not.
This is the Converse and Serve layer of a Smart Site, and it is the piece owners are most nervous about and most relieved by once it is running honestly. Built to these rules, it is not a replacement for your people. It is the reason they get to spend their time on the customers who need them.
Further reading
Sources & adjacent reading
Keep reading
Related Coffee Reads
What a Smart Website Is, and Why Your Brochure Site Is Costing You Customers
Most small-business websites are brochures: they describe, then go quiet. A smart website works the front desk. What that means in practice, the six things it does, what it is not, and how to tell which one you have.
The Internal Dashboard Every Small Business Should Have, and the Six Numbers on It
Owners run good businesses from memory, a group text, and six browser tabs. One screen with six numbers fixes most of that. What goes on it, where the data comes from, which automations to add first, and when to build versus buy.
The Chatbot Era is Dead: Why Big Tech is Scrambling for Secure AI Agents
The industry has pivoted from novelty chatbots to interconnected agentic workforces. Here is what the headlines mean for your business, and how to scale without becoming the next breach.