AI for customer support2 min read
AI customer service: automation and human review
What AI can automate
Automation works when the answer comes from a known source and the next step is already approved.
| Customer asks | The AI | Evidence it needs |
|---|---|---|
| “Are you open on Friday?” | Answers from your hours | Current opening hours |
| “Where is my order?” | Looks up the status | A matched customer and order |
| “Can I book tomorrow?” | Takes the booking request | Your services and booking rules |
| “Can you waive this charge?” | Passes it to your team | A person with authority |
Good first jobs:
- product details and opening hours;
- order and booking status;
- qualifying questions with set fields;
- appointment confirmations and reminders.
An AI reply agent should never invent a policy or promise a refund it cannot issue. Writing a reply and completing an action are separate jobs.
Chatbot or conversational AI?
A chatbot is the interface, and it may run on fixed menus with no AI at all. Conversational AI understands a request and replies across several turns. For either, ask what evidence it uses, what it may do and when it stops.
What the agent needs to know
| Question | Source |
|---|---|
| Policies, products, hours | Your knowledge base |
| A specific order, booking or account | Live records |
| What the customer already said | The conversation |
The conversation adds context, but it never overrides a policy. The grounding guide shows how to test whether a reply used the right source.
When a person should take over
- an exception to a refund, price or delivery policy;
- a complaint that disputes the facts;
- a safety, legal, financial or medical question;
- conflicting records, or a customer you cannot match;
- any commitment outside what the AI is allowed to do.
Set the boundary by the action and its impact, not by the model's confidence. A model can be confident about something it has no authority to do.
How Otobiz passes a chat to your team
Two things trigger it:
| Situation | What happens |
|---|---|
| The AI lacks the answer | It asks one question or says the team will check; a second miss goes to a person |
| A request for a person, a refund, a payment dispute, or a legal or medical question | Draft held; the customer is told your team will reply, and the AI stops |
In the shared inbox, the chat shows a Needs a person badge. A reply from your team takes over by default, and Take and Release let a person claim the chat or return it to the AI.
The customer stays in one thread and never repeats the request.
What to measure
Measure completed jobs, not messages sent:
- time to the first useful answer;
- conversations resolved without reopening;
- chats passed to a person, and why;
- failed actions, and whether anyone saw them;
- satisfaction on completed conversations.
Conversation analytics shows volume, response time and recurring topics. Revenue and refunds need your business records.
How to test a system
Run four scenarios with your own information:
- a routine question;
- a question your sources cannot answer;
- a customer-specific status check;
- a request that needs discretion.
For each, check the reply, its source, the action and where it stopped. The platform checklist covers ownership, consent and data export.
