AI assistants for customer support: what to automate and where a person must stay

AI assistants can now hold a reasonable conversation, understand questions written in everyday language, and answer from your own documents. For a business that receives the same questions every day through its website, Facebook page or WhatsApp, the appeal is obvious: instant answers at any hour, and fewer repetitive messages for the support team.

The risk is just as real. An assistant that answers confidently but wrongly, or that traps a frustrated customer in a loop, can do more damage than a slow reply. This article sets out what to automate, where a person must stay involved, and how to introduce an assistant safely.

A lesson worth remembering

In February 2024, a Canadian tribunal held Air Canada responsible for wrong information its website chatbot gave a customer about bereavement fares. As reported by CBC News, the airline argued the chatbot was responsible for its own actions. The tribunal rejected that, finding the airline responsible for all information on its website, whether from a static page or a chatbot, and that it had not taken reasonable care to ensure the chatbot was accurate.

Laws differ between countries, but the principle is sensible anywhere: customers will treat what your assistant says as what your business says. Design accordingly.

Your AI assistant speaks for your business. If you would not let a new employee say it unsupervised, do not let the assistant say it either.

What AI handles well

Good candidates for automation share three features: the questions are frequent, the answers are documented, and a wrong answer is easy to correct.

Factual questions from your own content: opening hours, locations, services offered, course schedules, delivery areas, required documents.
Status look-ups: "where is my order?", "is my appointment confirmed?", "has my payment been received?", answered from your systems after the customer is identified.
Guided steps: how to book, how to reset a password, how to request a return.
Collecting details before a hand-off: gathering the order number, the problem and a photo, so the person who picks up the conversation does not have to start again.
Sorting and routing: recognising whether a message is a sales enquiry, a complaint or a technical problem, and sending it to the right team.
Drafting replies for staff: suggesting an answer that an agent checks and sends, which saves time without removing judgement.

Where a person must stay

Some conversations should reach a human quickly, or be handled by one from the start:

Money decisions: refunds, compensation, discounts, exceptions to policy and anything that commits the business financially.
Complaints and upset customers: people who are angry or distressed want to be heard by a person.
Health, legal, safety or financial advice: a clinic's assistant can book an appointment, but should not interpret symptoms or results.
Account changes with security implications: changing a registered phone number, email or payout details.
Anything the assistant is unsure about: an honest "let me pass you to a colleague" is better than a confident guess.

Designing the assistant

Answer from your content, not from general knowledge

Configure the assistant to answer from an approved set of documents: your FAQs, policies, price lists and service descriptions. If the answer is not in those sources, it should say so and offer a person. This approach, often called retrieval, greatly reduces made-up answers, though it does not eliminate them, so testing still matters.

Keep the knowledge current

An assistant is only as accurate as its sources. Assign an owner for each document, and update the assistant's content whenever prices, policies or schedules change. Out-of-date policies are a common cause of wrong answers.

Make the hand-off easy and visible

Offer "talk to a person" clearly, at any point, not only after several failed attempts.
Pass the full conversation to the agent, so the customer does not repeat themselves.
Be honest about response times outside working hours.

Say it is an assistant

Tell customers they are talking to an automated assistant. People are generally comfortable with this when it is helpful, and uncomfortable when they discover it later.

Protect personal data

Verify identity before sharing order or account details. Do not ask for passwords, PINs or one-time codes. Check where conversation data is stored and processed, how long it is kept, and whether your AI provider uses it for training. In Bangladesh, the Personal Data Protection Act, 2026 sets obligations for handling personal data; take legal advice on how it applies to your assistant.

Measure what matters

Do not judge the assistant by how many conversations it handles alone. Track:

Resolution: how many customers got what they needed, confirmed by a short question at the end.
Accuracy: a weekly review of a sample of conversations by a senior team member.
Hand-off quality: whether agents receive enough context to help straight away.
Escalations and complaints: conversations where customers became frustrated or the assistant got something wrong.
Agent time: whether the team's repetitive workload has actually fallen.

A safe way to introduce it

Collect your top questions from recent messages and calls, and write clear, approved answers.
Start internally, with the assistant drafting replies that agents review before sending.
Go live on a narrow scope, such as opening hours, locations and order status, with an easy hand-off.
Review conversations weekly, fix wrong answers at the source, and widen the scope gradually.
Keep people in charge of refunds, complaints and anything sensitive.

Where RAED fits

We build AI assistants that answer from your approved content, connect to your order, booking or account systems, and hand conversations to your team with full context. We start with a narrow, well-tested scope and expand only when the results justify it.

Sources

CBC News, "Air Canada found liable for chatbot's bad advice on plane tickets", February 2024. cbc.ca
Securiti, overview of Bangladesh's Personal Data Protection Act, 2026, 29 June 2026. securiti.ai