Let a chatbot answer what your website already states in writing, such as opening hours, delivery areas and how to book. Have it draft, not send, anything about refunds, complaints, prices it can't confirm or personal data, and have a named person approve those. Tell customers it is AI, and give them a quick way to reach a person.
That is the decision this page is for. Most pages about an AI chatbot for small business compare tools; they rarely say which of your messages a bot should be allowed to finish on its own. This page sits under AI automation. For which task to automate first and what it costs, see is AI automation worth it for a small business. For how much access any AI helper should get to mail, files and CRM, see AI agent security.
What goes wrong when a chatbot answers customers on its own?
One failure is a confident wrong answer about policy or money, and the business still owns it. A tribunal in British Columbia made Air Canada pay a customer after its website chatbot described a refund rule that the airline's own policy page contradicted.
In Moffatt v. Air Canada, a passenger asked the airline's website chatbot about bereavement fares after a death in the family. The chatbot said they could book now and apply for the reduced fare within 90 days of the ticket being issued. Air Canada's own bereavement page said the policy does not cover requests made after travel, and the chatbot's reply even linked to that page.
According to the tribunal's decision, Air Canada argued it could not be held liable for what the chatbot said. The tribunal member called that "a remarkable submission" and wrote: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." The airline was ordered to pay the passenger $812.02 in damages, interest and tribunal fees.
The second failure is a bot that can be talked into saying anything. In January 2024 the parcel firm DPD switched off part of its online support chatbot after a customer got it to swear and to call DPD "the worst delivery firm in the world". DPD told the BBC: "An error occurred after a system update yesterday. The AI element was immediately disabled and is currently being updated." (BBC News).
Neither case needed a clever attacker. One was a wrong policy answer. The other was a customer typing requests into a support chat.
Do customers want a chatbot at all?
Many don't, and the main fear is losing the route to a person. In a Gartner survey of 5,728 customers run in December 2023, 64% said they would prefer companies didn't use AI for customer service, and 53% said they would consider switching to a competitor over it (Gartner, July 2024).
The top concern in that survey was that it would get harder to reach a person, ahead of jobs and wrong answers. Gartner's own advice in the same release is that AI chatbots "must communicate to the customer that they will connect them to an agent in the event that the AI cannot provide a solution", and then hand over so the person picks up where the bot left off.
Which messages can a chatbot answer, and which should a person approve?
Sort your messages by what a wrong answer would cost you. If the answer is already written on your site and a mistake is cheap, the bot can answer. If the reply promises money, a date, an exception or anything about a person's data, the AI drafts and a named person sends.
This is our rule of thumb, not a law or a vendor rule:
Message type | Bot answers | AI drafts, a person approves | Person only |
|---|---|---|---|
Opening hours, location, how to book, what a service includes (as written on your site) | Yes | ||
Order or booking status read from your own system | Yes, read-only | ||
Price quotes, discounts, anything not on your published price list | Yes | ||
Refunds, cancellations, exceptions to a policy | Yes | ||
Complaints and anything angry | Yes, or hand straight to a person | ||
Changes to an account, address or payment details | Yes | ||
Health, legal or safety questions, or anyone who seems distressed | Yes |
The Air Canada question belongs in the "AI drafts, a person approves" column: a policy exception with money attached. Under that rule, the chatbot's reply would have waited for someone who knew the bereavement policy, and that person could have stopped it.
What did our own inbox test show?
Our own test sorted messages without any AI model, which is a useful result in itself. On 16 and 17 October 2026 we built a local n8n workflow called Sort Inbox for a demo. It sorted 30 test emails into Lead 10, Support 7, Billing 7 and Spam 6 using keyword rules in a code step.
We published those figures in our AI automation cost guide and priced the same workflow both ways in n8n vs Zapier. They are test data on our own machine, not a client inbox.
Two things carry over to customer replies. Sorting a message into a tray is often cheap, predictable code, and you don't need a language model to do it. The model earns its cost at the next step, writing a first draft from your own documents, and that is exactly the step where a wrong sentence reaches a customer.
We did not have that test write or send replies, so we have no error rate of our own to quote for drafted answers. Treat any vendor accuracy figure the same way until you have run the bot against a week of your own real questions.
How should the approval step work?
The draft waits in an approvals queue, shows which document it came from, and goes out only after a named person clicks send. On our AI automation service that is the default: outbound messages wait for a person, and drafted replies cite the document they came from.
The person approving checks three things before sending:
- Does the cited document actually say what the draft says? If a draft has no source, it goes back.
- Does the reply promise money, a date or an exception? If yes, is that person allowed to promise it?
- Would you be comfortable seeing this reply in a screenshot on social media?
Name the approver before go-live, plus a stand-in for days off. A queue with no owner turns into a backlog that customers feel. If the bot also routes or logs messages unattended, watch what lands, not just the green status, as covered in automation monitoring for silent failures.
Can a customer trick the chatbot into doing something?
Yes, if the bot can do more than talk. Prompt injection is the first entry in OWASP's Top 10 for language-model apps, and its own example is "an attacker injects a prompt into a customer support chatbot, instructing it to ignore previous guidelines, query private data stores, and send emails" (OWASP LLM01:2025).
OWASP's mitigations include least-privilege access and "human approval for high-risk actions". In practice that means a customer-facing bot gets read-only access to what it needs to answer, and anything that writes, refunds or emails someone else goes through the approval queue. We treat text from forms and emails as untrusted input that must not steer the automation.
Do I have to tell customers they are talking to AI?
In the EU, yes, unless it is obvious. Article 50 of the AI Act requires providers to build AI systems that talk directly to people so that "the natural persons concerned are informed that they are interacting with an AI system", unless a reasonably well-informed person would already see it (AI Act Explorer, Article 50).
That duty sits with the provider, so if you install a vendor's chatbot, ask the vendor how it discloses itself. Outside the EU, the rules vary by country and state, so check the ones that apply to you. Disclosure is cheap either way, and the customer-facing bots we build say that they are AI.
Should I buy an off-the-shelf chatbot, use AI-drafted replies, or keep support human?
It depends on your volume and on how many of your questions are already answered in writing. For most small teams we would start with AI-drafted replies and a person sending, because the AI does the typing and the person keeps the judgement.
- An off-the-shelf chatbot suits a business with lots of repeat questions whose answers already sit on the site, and a clear "talk to a person" button.
- AI-drafted replies suit a business where most messages need a judgement call, such as quotes, bookings or anything with an exception.
- Keeping support human suits low volume, high-value customers, or messages that are mostly complaints and exceptions.
If you aren't sure which bucket you're in, export a month of customer messages and sort them into the table above before you buy anything.
FAQ
Will an AI chatbot replace my customer service?
For most small businesses it replaces some typing, not the person. Gartner found the top customer worry about AI in service was that it would get harder to reach a person, so a bot that removes the human route works against you. Use it for the repeat questions and keep a person on the exceptions.
Can a chatbot answer from my own FAQ and policy documents?
Yes, and it should answer only from them. Point it at pages you have already published, make every draft cite its source, and have it hand off when the answer isn't there. Air Canada's chatbot linked the right policy page and still contradicted it, so a person checks that the source says what the reply says.
What is Automation Lite, and can it handle customer replies?
Automation Lite is a permanent free offer. We set up one small automation, customised for you on a call, and you manage it on our platform at no cost. If that automation drafts customer replies, they wait for a person to approve them, which is the default for outbound messages on our AI automation service.
What should an AI chatbot never do without a person?
It should never change account or payment details, promise a refund or a policy exception, or answer health, legal or safety questions on its own. It also shouldn't email anyone other than the customer it is talking to. Those actions go to the approval queue or straight to a person.
What to do next
Export a month of your customer messages and sort them into the three columns above. If you would like help deciding which ones a bot can finish and which need a person, book a free 30-minute assessment.
How this guide was made: Ramin Moradi researched and wrote it with AI assistance. The tribunal decision, the BBC report, the Gartner release, the OWASP page and Article 50 were checked against the linked pages on 12 November 2026 (Europe/Istanbul). The Sort Inbox figures come from our own 16-17 October 2026 test and are the numbers already published on our cost guide and n8n vs Zapier post. The message table is our rule of thumb, not a legal standard.

Comments