Adding a website chat assistant: answers and handoff rules

Start from the questions you already get

Before choosing a tool, collect the questions people actually ask. Go through the last few months of contact-form messages, enquiry emails and notes from phone calls, and group them. Count how many fall into each group: opening hours, location, what is included, how to book, how long things take, and so on.

The biggest groups become the assistant's scope. Questions that appear once a quarter belong in the handoff pile.

What the assistant should answer

Good candidates share three properties: the answer is written down, it changes rarely, and getting it slightly wrong does no real harm. Examples are opening hours, directions and parking, what a service includes, how the booking or enquiry process works, which documents to bring, and where to find a policy.

Write each answer once, in a single file or sheet, and give it an owner and a review date. The assistant should answer only from that source, whether it uses fixed answers or an AI model limited to approved content. When a fact changes, you change one line, not a dozen pages and a chatbot.

What it should not answer

Keep these out of scope, and send them to a person or to the exact source page:

• Prices, availability and delivery dates, unless they are read live from the system that holds them. A copied price goes stale the day it changes.
• Anything contractual: refunds, cancellations, guarantees, eligibility. Link to the policy page and quote it word for word, rather than letting the assistant paraphrase.
• Advice that needs a professional: medical, legal, financial or safety questions.
• Anything about a specific customer's account, order or booking, unless the visitor is signed in and the assistant reads from your system with the right permissions.
• Complaints. People who are already unhappy should reach a person quickly.
• Questions unrelated to your business. A polite "I can only help with questions about [business]" is enough.

Why this matters: the Air Canada ruling

In Moffatt v. Air Canada (2024 BCCRT 149), a British Columbia tribunal dealt with an airline chatbot that told a customer he could apply for a bereavement fare discount after buying his ticket, which contradicted another page on the airline's own site. Air Canada argued, in the tribunal's words, that the chatbot was "a separate legal entity that is responsible for its own actions". The tribunal rejected that: "It should be obvious to Air Canada that it is responsible for all the information on its website." The airline had to compensate the customer.

The ruling is Canadian and from a small-claims tribunal, so it does not bind courts elsewhere. The practical lesson is general: treat every answer the assistant gives as something your business published.

Handoff rules

When to hand off

Write the triggers down. A workable starting set:

• the visitor asks for a person, in any wording;
• the question is on the "do not answer" list;
• the assistant could not answer twice in a row;
• the message mentions a complaint, cancellation, refund, urgency or health;
• the visitor shows buying intent that a person should handle, such as a group booking or a quote request.

What the person receives

The handoff should arrive with enough to reply without asking again: the visitor's name and contact details (with their consent), a one-line summary, the full transcript, the page they were on and the time. Send it where your team already works: a shared inbox, a WhatsApp chat or a CRM record, and alert the person on duty.

Out of hours

Say plainly when a person will reply, and make sure that promise is true. "We reply by 10:00 the next working day" is better than an instant-sounding promise the team cannot keep.

Saying it is an assistant

Tell visitors they are talking to an automated assistant, in the widget header or first message. In the EU this is becoming a legal duty: Article 50(1) of the AI Act requires that AI systems intended to interact directly with people are designed so that those people are informed they are interacting with an AI system, unless that is obvious from the context. The Act lists 2 August 2026 as the date this article applies; later EU amendments can change timing, so check the current text before you rely on a date. Outside the EU, the same line still avoids confusion and sets expectations about what the assistant can do.

Chat transcripts are personal data

Chats collect names, emails and phone numbers, and sometimes more sensitive information than you asked for. Under UK GDPR, the ICO's guidance on the right to be informed says people must be told your purposes, your retention periods and who their data will be shared with, at the time you collect it. So update your privacy notice to cover the chat, link it from the widget, decide how long transcripts are kept, and record which provider processes them, including any AI provider.

Test before launch

Build a test list from real questions, not ones written to pass. Include:

• the twenty or so most common questions, in the words customers used;
• questions on the "do not answer" list, which should hand off or link to a policy page;
• questions just outside scope, where a guess would be tempting;
• the same question in a second language, if you serve one;
• a request to "ignore your instructions" or to reveal internal information.

Record the expected outcome for each (answer, link or handoff) and the actual result. Launch when every handoff case hands off and every answer matches the approved source.

The scope sheet

Copy this into a shared doc and fill it in before the build starts.

WEBSITE CHAT ASSISTANT: SCOPE SHEET
Business:                     [name]
Approved answer source:       [link to sheet/doc]   Owner: [name]
Review every:                 [period]

ANSWERS (from the source only)
  [topic]                     [source row]
  [topic]                     [source row]

DO NOT ANSWER (hand off or link)
  Prices / availability       [live system or handoff]
  Refunds, cancellations      [link to policy page]
  Advice (medical/legal/etc.) [handoff]
  Account or booking details  [handoff]
  Complaints                  [handoff, priority]

HANDOFF TRIGGERS
  asks for a person | on do-not list | 2 failed answers |
  complaint/cancel/urgent/health | quote or group request

HANDOFF ROUTE
  Sends to:                   [inbox / WhatsApp / CRM]
  Includes:                   name, contact, summary, transcript, page, time
  Alert:                      [person or channel on duty]
  Out-of-hours promise:       [exact wording]

DISCLOSURE LINE               [e.g. "Automated assistant. A person can take over."]
PRIVACY                       notice link: [ ]  transcript retention: [ ]
                              processors: [widget, AI provider]

TEST LIST                     [link]   Passed on: [date]   By: [name]

The handoff route is where the assistant meets the inbox, CRM and alerts your team already uses, and it needs the most testing. That is the automation work we do: anything important reaches a person, with the details attached.

Sources

• Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia), as reported by McCarthy Tétrault, TechLex blog: https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot. Quotes: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions." and "It should be obvious to Air Canada that it is responsible for all the information on its website." Decision on CanLII: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html
• EU AI Act, Article 50(1) (transparency obligations), via the AI Act Explorer: https://artificialintelligenceact.eu/article/50/. The page states it applies from 2 August 2026 (Article 113).
• ICO, "Right to be informed": https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/right-to-be-informed/ (purposes, retention periods and recipients must be given at the time data is obtained; the ICO notes the guidance is under review following the Data (Use and Access) Act).

Related project and service