An AI chatbot for customer service works when it answers only from your approved content and live account data, resolves the simple, high-volume requests end to end (order status, appointment changes, policy questions, password and billing help) and hands everything else to a person quickly, with the conversation attached. It fails when it’s allowed to improvise, when it can’t see the customer’s actual account, or when it stands between customers and your staff.
Modern language models have made chatbots far more capable than the scripted decision trees of a few years ago. They’ve also made it easier to deploy one that sounds confident and is wrong. This guide covers how to build the first kind and avoid the second.
What can an AI chatbot for customer service do today?
Compared with older rule-based bots, an AI chatbot built on a large language model can understand questions phrased in many ways, in English and French, follow a multi-turn conversation and write natural replies. Connected to your systems, it can also act. Typical high-value tasks:
- Answer policy and product questions from your help centre, manuals, terms and internal knowledge base, citing the source.
- Look up account information such as order and shipping status, invoice balances, warranty coverage or appointment times, after verifying the customer.
- Complete simple transactions: reschedule an appointment, update a delivery address, resend an invoice, start a return.
- Triage and route: collect the details for a complex issue, classify it, and open a case in the right queue with a summary, so the agent doesn’t ask the customer to repeat themselves.
- Assist human agents: suggest replies, summarize long threads and surface the relevant policy inside the helpdesk.
The last item is often the best starting point. An assistant that drafts replies for agents to review delivers value quickly, builds the knowledge base and evaluation set you need, and carries little risk because a person sends every message.
Why do customer service chatbots fail?
They improvise
A general-purpose model asked about your refund policy will produce a plausible refund policy. If it isn’t yours, you have a problem. In Moffatt v. Air Canada (2024 BCCRT 149), the BC Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation after its website chatbot told a customer that a bereavement fare could be claimed retroactively, which contradicted the airline’s actual policy. The tribunal rejected the argument that the chatbot was a separate entity responsible for its own actions. The damages were modest; the principle is not. What your chatbot says, your company says.
They can’t see the account
Customers rarely ask general questions. They ask about their order, their bill, their appointment. A chatbot without access to your CRM, order system or booking platform can only say “please contact us,” which customers rightly see as a wall.
They trap people
64%
That number isn’t an argument against AI in service. It’s an argument against bad AI in service. A chatbot that resolves a simple request in thirty seconds and offers a person the moment the customer asks addresses exactly the concern customers raise.
Nobody measures quality
Teams launch, watch the “containment rate” rise and assume success. Meanwhile, customers who gave up are counted as resolved. Without reviewing real conversations against a quality standard, you don’t know whether the chatbot is helping or quietly hurting.
How should an AI customer service chatbot be built?
A reliable chatbot has five parts, each with a clear job.
| Component | What it does | Key design choices |
|---|---|---|
| Knowledge retrieval | Finds the relevant passages from approved content for each question | Which sources are approved, who owns them, how fast updates appear |
| Account tools | Reads and updates customer data through defined actions | Identity verification, least-privilege access, which actions need confirmation |
| Conversation model | Understands the question and writes the answer from retrieved content | Model chosen by evaluation, instructions to answer only from sources, tone in each language |
| Guardrails | Blocks out-of-scope topics, sensitive advice and unsafe actions | Topics list, escalation triggers, prompt-injection defences |
| Handoff | Passes the conversation, customer identity and summary to a human | Channel (chat, ticket, callback), business hours, queue routing |
Answer from sources, and show them
The chatbot retrieves passages from your help centre and policies and answers only from them, linking to the source page. If it can’t find a supported answer, it says so and offers a person. This approach, called retrieval-augmented generation, is the same one behind a good AI assistant on company knowledge, and it makes your content team part of the chatbot’s quality.
Connect to live data carefully
Account lookups and transactions go through specific, tested actions that call your CRM, ERP, billing or booking system with scoped permissions, never through broad database access. Verify the customer’s identity before revealing or changing anything, and ask for explicit confirmation before transactions. Our guide to connecting AI to CRM and ERP systems covers the integration patterns in detail.
Design the handoff as carefully as the bot
The customer should be able to reach a person at any point by asking. When the chatbot hands off, the agent receives the full conversation, the verified identity and a short summary, and the customer is told what happens next and how long it will take. Outside business hours, offer a callback or a ticket instead of a dead end.
Which requests should you automate first?
Pull three months of tickets, chats and call reasons and sort them by volume. The best first candidates are:
- High volume. The top five to ten reasons usually account for a large share of contacts.
- Answerable from data or documented policy. Order status, opening hours, how-to questions, document requests.
- Low judgement. No exceptions, negotiation or empathy-heavy situations.
- Low risk if wrong. An incorrect store hour is an annoyance; an incorrect cancellation fee is a complaint.
Keep complaints, cancellations with retention offers, billing disputes, anything involving vulnerable customers and anything with legal or safety implications with people, at least at first.
What about privacy and language requirements in Canada?
Customer conversations contain personal information, so the usual Canadian obligations apply:
- PIPEDA governs how private-sector organizations collect, use and disclose personal information in commercial activity, and Québec’s Law 25 adds requirements such as privacy impact assessments for certain projects and transparency about automated decisions.
- Tell customers they are talking to an AI assistant, and explain how conversations are used.
- Keep transcripts in Canadian regions where possible, use model endpoints that don’t train on your data, set retention periods and restrict who can read transcripts.
- Mask payment card numbers and other sensitive identifiers before they’re stored or sent to a model.
- For Québec customers, service in French isn’t optional. Test the chatbot’s French with native speakers on your real questions, not just translated test cases.
Our security practices describe how we handle data residency, access and audit by default.
How do you measure whether the chatbot works?
Track a small set of measures from day one, and review a sample of real conversations every week:
- Resolution rate: conversations where the customer’s need was actually met, confirmed by outcome data or follow-up, not just conversations that ended.
- Answer accuracy: percentage of reviewed answers that are correct and supported by an approved source.
- Handoff quality: time to reach a person, and whether agents had to ask customers to repeat information.
- Customer satisfaction for chatbot-only and handed-off conversations, compared with your existing channels.
- Agent time saved and changes in queue wait times.
- Cost per resolved conversation, including model usage.
Our guide to measuring AI ROI shows how to turn these into a business case your CFO will accept.
A realistic rollout plan
| Phase | Weeks | What happens |
|---|---|---|
| Discover | 1–2 | Analyse contact reasons, pick the first request types, audit the knowledge base, agree on metrics |
| Prototype | 3–5 | Agent-assist or internal chatbot on real questions, measured against a test set of past conversations |
| Build | 6–10 | Integrations, identity verification, guardrails, handoff, French and English testing, privacy review |
| Launch and improve | 10+ | Soft launch to a share of traffic, weekly conversation reviews, gradual expansion of request types |
How we build customer service assistants
Our AI assistants answer from your approved content with citations, in English and French, and our AI agents handle the account lookups and simple transactions through your systems with scoped permissions and full audit logs. Everything runs in your own cloud account in a Canadian region, models are chosen by evaluation on your real conversations, and a person is always one request away.
To see what an assistant for your contact centre would involve, describe your top request types in our free AI blueprint for a solution outline and timeline in a couple of minutes, or book a call with a senior engineer.
