Picture a Saturday night in November. A customer emails to ask whether the patio heater she ordered will arrive before her party on Friday. Nobody will read the email until Monday.
That is the moment most owners have in mind when they ask whether AI should handle customer service. The answer is yes, but not in the job most of them first imagine. Let AI answer the questions your business has already answered in writing, and let it draft the replies your team sends. Keep a person on anything that makes a promise, involves an upset customer or touches money. Customers are clear that they want a route to a human, and one tribunal has already made the other point plain: whatever your chatbot says, your business said.
Where does AI actually help in small business customer service?
AI helps most with the reading and writing in customer service, and least with the deciding.
Look at what fills a customer inbox in a typical small business. Most of it is reading: working out what the customer wants, which order they mean, whether this is the third message about the same problem. Then comes writing: a clear, polite reply that gets the facts right. Deciding is a thin slice at the end, and it is the slice that matters most.
AI is good at the first two. Inside the tools many businesses already pay for, such as Microsoft Copilot in Outlook, Gemini in Gmail or ChatGPT, a member of staff can:
- sort the morning's messages into orders, returns, complaints and questions;
- summarize a long email thread before replying to it;
- draft a reply from your own written policies, in your house tone;
- turn a scrawled phone note into a tidy order update;
- rewrite a help page so a customer can find the answer before they need to ask.
None of these puts AI in front of a customer on its own. That is deliberate, and it is where I would start.
Most small firms are still near the beginning. The UK's Office for National Statistics reported on July 20, 2026 that around 35% of businesses with ten or more employees now use at least one AI technology, and 28% of businesses with fewer than ten. The same release found that improving business operations is the most common use. Customer service sits right on that line: an operation behind the scenes, with a customer at the end of it.
What do customers want from AI customer service?
Customers welcome AI that makes contact easier and resent AI that stands between them and a person.
A Gartner survey of 3,566 business and consumer customers, published on August 4, 2026, found that 50% of customers say their interactions are easier when companies use generative AI. It also found that 87% say it is essential for a company using it to offer a way to reach a human agent. Eric Keller, a senior director analyst at Gartner, put the conclusion in one line: "Service leaders should not use GenAI as a mandatory first step for every issue."
Two further findings from the same survey matter more to a small business than the headline. Customers were about three times more likely to use their own AI tools, such as ChatGPT, Gemini or Copilot, than a company's chatbot. And 58% of customers who use generative AI have used it to complete a task on their behalf, rising to 74% in business-to-business settings.
My reading: your customers are already bringing AI to you. Some will have asked ChatGPT about your returns policy before they email. Some of the emails you receive were drafted by AI. That changes the first job. Before you build a chatbot, make sure the answers it would give are written down clearly on your own site, because machines as well as people are now reading them.
Who is responsible when an AI chatbot gets it wrong?
Your business is, in the same way it is responsible for anything else on its website.
The case owners should know is Moffatt v. Air Canada. A customer asked the airline's website chatbot about bereavement fares and relied on its wrong answer. On February 14, 2024, British Columbia's Civil Resolution Tribunal ordered the airline to compensate him. Air Canada had argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that and held the airline responsible for all the information on its website, whether it came from a static page or a chatbot.
It is one Canadian decision, not a rule in the UK or the US, and the sum involved was small. But the reasoning is ordinary common sense, and it is the reasoning I would plan around. A chatbot does not have its own opinions about your refund policy. It has yours, or it has something it made up that customers will treat as yours.
The Promise Rule: what AI may say to your customers
The Promise Rule is my suggested way of deciding what AI may say to a customer, and it sorts by promise rather than by topic.
The Promise Rule is the line I suggest a small business draws before AI speaks to a single customer: AI may tell a customer anything the business has already written down, but it may never promise anything a person has not already decided. Questions with a written answer can go to AI. Replies that make a promise are drafted by AI and sent by a person. Anything involving an upset customer, money outside policy or a problem nobody has seen before goes straight to a person.
Here is how it sorts a week of messages in a hypothetical 15-person online garden supplies retailer heading into its busiest quarter.
| Message | Written answer exists? | Makes a promise? | Lane |
|---|---|---|---|
| "What time does your phone line open?" | Yes | No | AI can answer |
| "Where is my order?" | Yes, the tracking record | No | AI can answer |
| "Can I return this after 40 days?" (policy says 30) | No, it needs an exception | Yes | AI drafts, a person decides and sends |
| "Will it arrive by Friday?" | Partly | Yes, a delivery date | AI drafts, a person sends |
| "Second broken planter this month. Really disappointed." | No | Whatever you say next is one | A person, from the first word |
Go back to the patio heater. Under the Promise Rule, AI can tell the customer on Saturday night what the tracking record says, because that is written down. It cannot tell her the heater will arrive by Friday, because nobody has decided that. The honest reply says what the record shows and when a person will confirm. It is less satisfying than "Yes, no problem!", and far cheaper than a promise you cannot keep.
Notice what the rule does not ask. It does not ask whether the AI is clever enough. Modern tools can write a fluent answer to every row in that table. The rule asks whether a person in your business has already made the decision the answer contains.
The Mistake I See Most Often
The mistake I see most often is starting with the chatbot on the website, the most visible and riskiest place, instead of the inbox behind it, where a person still reads every word.
When I work with business teams on customer service, the conversation almost always opens with the widget in the corner of the homepage. It is what the owner has seen elsewhere and what vendors show. But the website chatbot is the one place where AI talks to customers with nobody watching. The inbox is the opposite: AI drafts, a person reads, the customer only sees what the person approves. Every mistake is caught, and every draft teaches the team something about what good looks like.
That last point comes from my first career. I spent two decades as a teacher, assistant headteacher and Director of Digital Strategy in further education before I worked with businesses, and the most reliable way I know to raise the quality of anyone's writing is to let them edit good examples. An AI draft that a skilled customer service person corrects every morning is a training program in disguise. As I put it on my business page, the hard part of AI at work isn't the technology. It's helping people use it well, safely and every day.
The principle underneath is the one I teach everywhere: outsource the doing, not the thinking. Drafting a reply is doing. Deciding what your business will promise is thinking.
What should a small business do this week?
This week, sort your real messages, write down your answers and start AI in the inbox before it goes anywhere near a website. A 20-person business can do this in a few hours.
- Pull your last 50 customer messages. Emails, web forms, social messages, whatever you have. Sort each into the three lanes of the Promise Rule. Most owners are surprised how many sit in the first lane.
- Write down the answers to your ten most common first-lane questions in one short document. This becomes your single source of truth. If an answer is not written down, AI does not say it.
- Give AI drafting to the person who already answers the inbox. For two weeks, they draft second-lane replies with the AI tool you already pay for and send nothing unedited. Keep payment details and anything sensitive out of the prompt, in line with your one-page AI policy.
- Add one line to that policy: "AI may not promise a customer anything a person has not already decided."
- Put the Exit Question to the trial before it starts. The Exit Question is the one question I put to a leadership team before any AI pilot starts: if this works, what will change, who has already agreed to it, and on what date will you decide? A pilot without an answer is not a pilot. It is an experiment with no exit.
- Only then consider a customer-facing chatbot, limited to first-lane questions, fed only from your written answers, with a visible route to a person on every screen.
If this feels slow, it is the same logic as starting with one job rather than one tool. Customer service is simply the job where the customer sees your mistakes.
When should AI stay out of customer service?
Keep AI away from your customers entirely if your answers are not yet written down, if a handful of large clients make up most of your revenue, or if customers come to you for regulated advice.
The first case is the most common. A chatbot with nothing to read will improvise, and improvisation is exactly what the Promise Rule forbids. Write first. The second case is a business-to-business firm where every relationship is personal and each account is worth a year of effort: there, AI belongs behind the scenes, drafting and summarizing, never answering. The third covers financial, legal, medical and similar advice, where a wrong answer is more than an apology.
The counterargument deserves its due. A one-person business drowning in the same five questions has the most to gain from answering them automatically and the least time to wait. For that owner, a small first-lane chatbot, fed from a clear written page, may be the right first move. The rule still holds: it answers what is written, and it hands everything else to you.
Owners who skip these steps often end up with AI that everyone pays for and nobody trusts, the pattern I describe in why most small businesses are not getting value from AI yet. Customer service is too close to your reputation to let that happen.
Where to go from here
If your team already uses AI to answer customers but nobody has shown them where the line sits, that is exactly what my AI training for teams is built for. If you would rather find out where your business stands first, take the free Workplace AI Readiness Check: 21 questions, about eight minutes.
Sources and further reading
- Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent, Gartner, August 4, 2026.
- BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot, Lisa R. Lifshitz and Roland Hung, Business Law Today, American Bar Association, February 2024 (on Moffatt v. Air Canada, decided February 14, 2024).
- Artificial intelligence in UK businesses: 2023 to 2026, Office for National Statistics, July 20, 2026.
Dan Fitzpatrick helps businesses use AI well through keynotes, practical AI training for teams, and AI strategy and governance for leadership teams.
