AI & Automation · 10 min read
AI for Small Business, Part 4: AI for Customers, Answers and Onboarding
Part 4 of a weekly series: use AI for customer answers and onboarding with an approved library, three levels of trust and four numbers that show whether it works.

A customer writes to a small bakery at 9:47 on a Saturday night: "Do you do gluten-free birthday cakes, and can I pick one up tomorrow at 8?" The owner is asleep, as owners should be at 9:47 on a Saturday. By Sunday at 8, the customer has bought a cake from the place with a bot that answered.
Now picture the opposite. The bot answered at 9:47 with great confidence and told the customer, "Yes, gluten-free, ready at 8, 20% off." None of that was true. The bakery does not make gluten-free cakes, opens at 9 on Sundays and has never discounted anything in its life. The bot was not lying. It was doing what these tools do: producing a fluent answer, with no idea whether it is correct.
This is Part 4 of AI for small business, step by step. Part 1 drew the map, Part 2 handled follow-up and Part 3 took on the paperwork. Today's decision: how can AI help you answer customers fast and welcome new ones smoothly, without ever making a promise you cannot keep?
The one rule: the answer comes from your box, not from the bot's imagination
Think of the bakery's recipe box from the cover picture. Every card is an answer a person wrote, checked and approved: opening hours, allergens, how far ahead to order, the cancellation rule, the delivery area. The card is the fact. The assistant's job is to find the right card and write a friendly reply around it. It should never be the author of the card.
That gives you three levels of trust, and most small businesses should live at the first two.
- Level 1, AI drafts and a person sends. The assistant reads the customer's message, finds the matching approved answer and writes a reply. You read it, change a word and send. Time saved: most of the typing. Risk: very low, because a person is the last step.
- Level 2, AI answers alone, but only from the library. The assistant can answer only questions that match an approved card, quotes it, and says "let me get a person" for everything else. Risk: low, as long as "everything else" really is handed over.
- Level 3, AI answers anything it likes. This is the 9:47 bakery story. For a small business this level has almost no upside and a very large downside.
| Fine to automate | Keep a person in charge |
|---|---|
| Opening hours and allergens | Price quotes for custom work |
| How far ahead to order | Refunds and cancellations |
| Where is my order status | Deadlines you have not confirmed |
| Welcome message and reminders | Safety, health and legal questions |
A running example: Cedar Street Bike Repair (invented)
To make this concrete, meet Cedar Street Bike Repair, an invented shop with two mechanics and an owner named Marta who answers email at lunch and after the kids go to bed. The shop, the numbers and the customers in this post are illustrative, not a real business.
Marta's inbox has three kinds of messages. About half are "are you open, how long does a tune-up take, do you fix e-bikes?" About a third are "where is my bike?" The rest are real conversations: a quote for a rebuild, a complaint, a warranty question. She writes the same five answers all week, so she begins with the five cards.
Job 1: Answers
What AI is good at here. Matching a messy question to the right card, even when the customer wrote it in a hurry with no punctuation. Rewriting a card in a warmer or shorter tone. Answering in another language, with a person who can read the language checking the first few. Summarizing a long thread so that whoever picks it up does not start from zero.
Where it goes wrong. It fills gaps. If the card says "tune-ups take two to three days" and the customer says "it's urgent," a free-running assistant may invent a rush service. It also repeats old information: a price changed in October and the bot still quotes September. And it can sound sure of itself when it is guessing.
How to use it safely.
- Write the cards first. Five to twenty is enough to start. Each card has the answer, the date it was last checked and the name of the person who owns it.
- Tell the assistant, in writing, what it may use: "Answer only from the cards below. If no card fits, say you will ask a person and do not guess."
- Keep a "never answer" list: price quotes for custom work, refunds, deadlines you have not confirmed, anything about safety, health, legal rights or someone's account.
- Make handoff easy. Every automated reply must show how to reach a person, and a customer who writes "human" or "call me" gets one.
- Say that it is an assistant. Customers forgive a bot; they do not forgive a bot that pretends to be Marta.
A prompt that works. Paste your cards, then:
You are the assistant for [shop name]. Reply to the customer message below using only the approved answers above. Keep it to four sentences, friendly and plain. If no approved answer fits, or the customer asks about prices for custom work, refunds, safety or a deadline I have not confirmed, do not answer: write "Thanks, I'm passing this to [name], who will reply by [time]" and list what the customer asked, in one line, for me. Never invent a price, a date or a policy.
The story that every owner should know
In February 2024, a tribunal in British Columbia decided a dispute between a customer and Air Canada. According to the law-firm summaries I read, the customer asked the airline's website chatbot about bereavement fares, and the chatbot said he could apply for the reduced fare after travelling. A linked policy page said the opposite. The tribunal found the airline liable for negligent misrepresentation and rejected the argument that the chatbot was a separate party responsible for its own words.
I could not open the decision itself in this run, so treat the summary as secondhand, check the published decision before you quote it, and note that it is a small-claims tribunal ruling from Canada, not a rule that binds a Utah court. Even so, the lesson for a small business is practical and does not depend on the details: the customer does not care which part of your website said it. If your tool said it, you said it. That is why the library, the "never answer" list and the handoff matter.
Utah has its own rules about telling customers when they are talking to generative AI. The state's Artificial Intelligence Policy Act took effect in 2024 and was amended in 2025, and law-firm summaries describe the rules as depending on the kind of interaction. I did not read the statute in this run. Confirm the current text on the Utah Legislature's website, or ask a lawyer, before you put a customer-facing assistant live. A plain "I'm an assistant" at the start of a chat is good practice anyway.
Job 2: Onboarding
Onboarding is what happens between "yes, I'll buy" and "it works for me." It is also a process that most small businesses have never drawn, which is why the first week with a new customer often feels like improvisation. It is the same ground as the process teardown of the handoff, and it belongs to operations: if you do it well, a customer's first week feels like the shop has done this before.
What AI is good at here. Writing the welcome message from your template and the customer's details. Turning your messy notes into a short "what happens next" page. Drafting reminders. Producing a first-week checklist per customer type. Summarizing an intake form for the person who will do the work.
Where it goes wrong. It can promise a start date you did not choose. It can paste the wrong customer's detail into a message when the data is messy. It can write a perfect-sounding welcome that no person has ever read, so the customer's first real contact is a stranger.
How to use it well.
- Draw the first week on one page: day 0 (welcome and what we need), day 1 (a person calls or writes), day 3 (first milestone), day 7 (a check-in). Dates and owners come from you.
- Put the facts in a small intake form: name, what they bought, the one thing that matters most to them, the contact they prefer. The assistant reads the form and does not guess.
- Let automation send the welcome, the reminders and the check-in, and keep the first human hello for a person, even if it is two lines.
- Add one exit: "Reply HELP and a person will call you."

A prompt that works.
Write a welcome email for a new customer of [shop name]. Use only the details in the form below and the first-week plan below. Four short paragraphs: thank them, say in one sentence what happens next, list the two things we need from them with a date, and tell them who their person is and how to reach them. Do not promise anything that is not in the plan. If a detail is missing, write [MISSING] instead of guessing.
Guardrails that apply to every customer-facing use
- A person owns it. Name one owner and one backup. If nobody owns the cards, they go stale.
- Disclose. Say it is an assistant, at the start.
- Hand over fast. Any sign of anger, money, safety or confusion goes to a person within a time you promise and keep.
- Protect the data. Do not paste customers' payment, health or account data into a tool unless you have checked what the tool does with it. The one-page AI policy has a traffic-light list for this.
- Read a sample every week. Read at least five replies every week, and mark the wrong ones. This is how a library gets better.
- Have a fallback. If the tool is down, a person answers from the same cards.

Four numbers that tell you if it works
Do not judge this by how clever the replies sound. Judge it by numbers you can count in a week, the same way the Project Focus scoreboard does.
- Time to first reply. Hours between the customer's message and the first useful answer. Measure it before and after.
- Answers corrected by a person. Of the replies you read, how many needed a change? If it is zero, you may not be reading closely enough. If it is most of them, the cards are not good enough yet.
- Handoffs to a human. How many conversations the assistant passed on, and how many of those should have been passed. Too few is a danger sign.
- Customers who finish onboarding on time. The share who completed the first-week steps by day 7. It tells you whether onboarding is a process or a hope.
For Cedar Street, say Marta writes the number down for four weeks before she changes anything. In week 5, with the cards and the draft prompt in place, she measures again. If the first-reply time drops and the corrected-answer count falls week after week, she moves a few cards to level 2. If not, she stays at level 1. Either result is useful. These numbers are illustrative.
The hall of fame of customer-answer mistakes
- The Confident Discount. The bot offers a promotion that never existed. Fix: a "never answer" list that includes prices and offers.
- The Wrong Month. The cards were last checked in spring. Fix: a "last checked" date on every card, and a monthly review.
- The Mystery Bot. A customer asks whether they are talking to a person and gets a coy answer. Fix: say it is an assistant at the start.
- The Locked Door. No way to reach a human. Fix: an exit in every message.
- The Welcome Nobody Wrote. An automated welcome with another customer's name in it. Fix: [MISSING] markers, a test on three fake customers and a person reading the first ten.
- The Silent Fallback. The tool goes down and nobody knows how to answer by hand. Fix: the cards on one page, printed.
A 30-day plan
Pick your start date below and the tool lays out four weeks: write the cards, list what the assistant may not answer, test on old messages, run at level 1 with a person sending, measure, then decide whether to move anything to level 2. Move any date to fit your life.
What to do this week
- Open your inbox and list the ten questions you answered most often in the last month. Those are your first cards.
- Write each answer once, with a "last checked" date and an owner.
- Write the "never answer" list: price quotes, refunds, deadlines, safety, health, legal, anything about someone's account.
- Time your first reply today, and write it down. It is your baseline.
- Try the answer prompt on five old messages and compare the draft with what you really sent.
Next in the series is Part 5, on measuring the return: hours saved, errors avoided, revenue won. If you want help choosing the first workflow in your own business, or setting up the library and the checks that keep a person in charge, get in touch, read about AI and automation consulting for small business or start with the free templates in the toolkit. For the money side, measuring the return of automation comes first.
AI tools change quickly and make mistakes. The bakery, the bike shop and their figures are invented for the example. The Air Canada case is summarized from published law-firm commentary and should be checked against the decision itself; the summary of Utah's rules is not legal advice, so confirm the current law with the Utah Legislature or a qualified professional.
Frequently asked questions
Can I put an AI chatbot on my website without risk?
Not without risk, but you can keep it small. Limit it to approved answers, tell visitors it is an assistant, give an easy way to reach a person and read a sample of its replies every week. In a 2024 British Columbia tribunal decision about an airline's chatbot, as reported by law firms, the company was held responsible for what its chatbot said.
Do I have to tell customers they are talking to AI?
Disclose it as a habit. In Utah, rules on disclosing generative AI to consumers exist and were amended in 2025, and what they require depends on the kind of interaction, so confirm the current text on the Utah Legislature's website or ask a lawyer. Elsewhere, other rules may apply.
What should the first AI task for customers be?
Drafting replies to your most common questions, with you sending them. It saves typing, keeps a person in charge and shows you quickly which of your answers are unclear or out of date.
Should onboarding be automated?
Automate the sending, the reminders and the checklists, and keep the first human hello. Draw the first week on one page before you build anything, because automating an unclear process only makes the confusion faster.
Everything new, once a month.
One email on the 1st of every month with everything new: articles, series parts, news, website updates, project results, where I am working and the business opportunities I am looking at. No spam.
- Fresh on the site right now
- ArticleMeasuring the Return of Automation: Hours, Errors and Cash, Counted Honestly
- NewsSix North Temple murals, $150,000: what a project manager sees in the city's mural program
- SeriesOperations that scale from 3 to 30 people
Join the conversation
Did this help? A quick reaction tells me what to write next.
Be the first to react.

Loading comments…