To implement an AI chatbot for customer service, start with the questions you answer most often, give the AI accurate source material to answer from, and decide exactly when a conversation goes to a person. Then test with your own team, launch on one or two channels, and review the conversations every week. Most of the work is in the preparation, not the technology.
This guide covers each step in order, with the decisions you'll need to make along the way.
What "AI chatbot" means in 2026
Most people still say "chatbot", but there are two different things under that word.
A traditional chatbot follows a script. It matches keywords or button choices to pre-written answers. It's predictable, which is useful for fixed journeys like booking a slot or choosing a department, but it breaks when a customer phrases something in a way you didn't plan for.
An AI agent reads the customer's message, works out what they're asking and writes an answer from the knowledge you've given it. It copes with typos, slang and questions phrased in unexpected ways. When it can't answer, it should hand over to a person.
Most teams use both. A scripted flow handles structured tasks, and an AI agent handles open questions. If you want the longer version, read chatbot vs conversational AI.
Interest is high. A Gartner survey of 187 customer service leaders found 85% planned to explore or pilot customer-facing conversational GenAI in 2025. The teams that get value from it are the ones that prepare properly.
Step 1: decide what the AI should handle
Pull the last 30 to 90 days of conversations and sort them by topic. You're looking for questions that are:
Frequent. They come up every day.
Answerable from information you already hold. Opening hours, delivery times, returns policy, how to book, what documents to bring.
Low risk if the answer needs a follow-up. A wrong delivery estimate is fixable. Wrong advice on a contract or a medical question isn't.
Then write down what the AI must not answer. Typical examples are complaints, cancellations, anything legal or financial, and anything involving a vulnerable customer. These go straight to a person.
Start small. Five to ten well-defined topics is enough for a first launch.
Step 2: prepare your knowledge
An AI agent is only as good as what it's trained on. If your help articles are out of date or contradict each other, the AI will repeat the mistakes.
Clean up your source material
Remove outdated policies and old price lists.
Resolve contradictions. If the website says 14 days for returns and an internal doc says 30, decide which is right.
Write down the unwritten rules your team follows. These gaps are often behind the questions that cause the most back-and-forth.
Check the tone. If your documents are dry and your brand is warm, the AI will sound out of place.
Structure it so it can be found
Break long documents into short, focused sections, each answering one question. Use clear headings. A page titled "Returns" that covers returns, exchanges, refunds and warranty claims is harder to use than four short pages.
In Cue, you train AI Agents by adding documents (pricing, products, returns, terms and so on), connecting your website content and giving the agent custom instructions on name, tone and what to avoid. Our guide to building a knowledge base your AI can use goes further.
Step 3: choose your channels
Put the AI where your customers already contact you. For many UK and South African businesses, that means WhatsApp and web chat first.
WhatsApp has a few advantages for service. More than 3 billion people use it, the conversation stays in one thread on the customer's phone, and it supports buttons, images, documents and location sharing. Web chat catches people while they're still on your site, mid-decision.
Cue's AI Agents work on WhatsApp, web chat, email and Facebook Messenger. They don't answer phone calls. Calls in Cue are answered by your team.
Bring channels into one inbox
Customers move between channels, and they don't want to repeat themselves. Cue's omnichannel inbox brings WhatsApp, phone, email, web chat and Facebook Messenger into one inbox. If a conversation starts on Messenger and moves to WhatsApp, the history follows.
That matters for handover. When the AI passes a conversation to a person, or a customer who chatted yesterday rings today, your agent sees what has already happened. In Cue, calls land in the same inbox as messages, matched to the contact by phone number, with recordings and transcripts on the contact record. See how calls work in Cue.
Step 4: design the handover to people
Handover is where most AI deployments succeed or fail. Customers forgive an AI that says "I'll pass you to the team". They don't forgive one that keeps answering the wrong question.
Decide your triggers:
The customer asks for a person.
The topic is on your "never answer" list.
The AI can't find an answer in its knowledge.
The customer has repeated themselves or sounds frustrated.
The enquiry is high value, such as a large order or a sales lead.
Then decide where each conversation goes. In Cue, you build these rules in Flows, the no-code builder, and route handovers to the right team. The agent who picks it up sees the full chat, so the customer doesn't start again.
Plan for out of hours too. If no one is available, tell the customer when they'll hear back, and make sure the conversation is waiting at the top of the queue in the morning.
Step 5: test before you launch
Don't switch the AI on for every customer on day one.
Test internally. Have your team ask real questions from your conversation history, including awkward ones. Note every wrong or weak answer and fix the source material.
Launch on a slice of traffic. One channel, one brand or one site. Two weeks is usually enough to see the edge cases: regional phrasing, product names customers get wrong, questions nobody predicted.
Review every handover. Each conversation the AI passed on tells you something. Either the knowledge has a gap or the topic should stay with people.
Tell customers they're talking to AI
Be upfront. Name the agent, say it's AI, and make it easy to reach a person. It's good practice, and the EU AI Act includes transparency rules on telling people when they're interacting with an AI system. If you serve EU customers, check how those rules apply to you. For personal data, your obligations under UK GDPR, GDPR and South Africa's POPIA still apply to AI conversations as they do to any other.
Step 6: measure what's working
Track a small set of numbers from the start, so you have a baseline:
Metric | What it tells you |
|---|---|
Resolution rate | Share of conversations the AI closes without a person |
Handover rate and reasons | Where the knowledge has gaps, or where topics should stay with people |
First response time | How quickly customers get a first reply, by channel |
Customer satisfaction | Whether faster answers are also good answers |
Cost per resolved enquiry | AI resolutions compared with your team's cost to handle the same enquiry |
On cost, use your own numbers. Work out what a conversation handled by your team costs (agent time multiplied by hourly cost, including salary on-costs, is a fair start), and compare it with the AI cost. On Cue, an AI-resolved enquiry costs £0.99 in a bundle, and the Automate plan includes 500 AI resolutions a month. A resolution is a customer enquiry your AI agent closes without a person picking it up. See Cue pricing for the detail.
For a real example: Payflex, a buy now, pay later provider in South Africa, resolves 82% of customer queries with Cue's AI on its digital channels, with complex queries routed to live agents. Your rate will depend on how repetitive your queries are and how good your source material is.
For more on which numbers to track, see our list of customer service metrics.
Step 7: keep improving
Implementation isn't a one-off project. Set a weekly review for the first few months:
Read a sample of AI conversations, good and bad.
Close the knowledge gaps behind the most common handovers.
Add one or two new topics once the first set is working.
Update the source material whenever a policy, price or product changes.
Once the basics are solid, you can connect the AI to other systems so it can look up order status or booking details, and use it for lead qualification and appointment booking as well as support. Cue's integrations cover CRMs, dealer management systems and other business tools.
Ready to plan your rollout?
If you want to see how AI Agents, handover and calls work together in one inbox, book a demo and we'll use your own questions and channels.
