The key points
- Use three tiers: fully automatic, AI drafts and a person approves, and human only.
- Base the split on your real ticket history, not on a vendor's headline percentage.
- Anything involving money, safety, legal risk or an upset customer stays with a person.
- WhatsApp's policy expects a clear route to a human when you automate replies.
Which support questions are safe to automate fully?
The safe tier is questions with one correct answer that can be read from a system or a document: where is my order, what are your opening hours, how do I reset my password, what does the plan include. A chatbot connected to your systems can answer these at any hour, and a wrong answer costs little because the customer can see it is wrong.
The test is whether the answer can be checked against a source. If the bot looks up the order status from your system, it is reliable. If it improvises a policy from memory, it is not. Give the bot only the sources it should answer from, and make it say when it does not know.
What should AI draft but a person approve?
The middle tier is where most of the time is saved and most of the risk sits. Questions about a specific situation, such as a late delivery, a billing mistake or a product that does not fit, need judgement and tone. Let the AI read the history, pull the facts and draft a reply, then have an agent edit and send it. The agent does a few seconds of checking instead of several minutes of writing.
Over time, look at which drafts are sent unchanged. Those categories can be promoted to fully automatic, with a spot-check. Categories where agents keep rewriting stay in the middle tier. This is the practical route to the scope of an AI automation project: start narrow, measure, widen.
What must stay with a human?
Keep refunds above a limit you set, complaints, cancellations of long contracts, anything legal or safety related, and any message where the customer is clearly upset. Also keep a human route open everywhere else. WhatsApp's business policy allows automation inside the 24-hour window but expects prompt, clear and direct escalation paths, such as handing the chat to a person or giving a phone number or email.
Customer messages contain personal data, so decide what the AI may see and where it is stored. Our note on how your data is handled when you add AI features covers the questions to ask.
Before you automate support
- A few hundred past tickets are sorted into automatic, draft and human tiers
- Automatic answers come from a system or approved document, not from memory
- Drafts are reviewed by a person before they are sent
- Refunds, complaints and sensitive cases route to a human
- Every automated chat offers a clear way to reach a person
- Someone reviews wrong or rejected answers every week
Sources
Rules and guidance this answer relies on. Policies change, so check the current version before acting on it.