What an AI agent should and should not answer on your business WhatsApp

The argument about AI on customer messaging usually gets framed as all or nothing: either you automate replies or you keep answering everything by hand. Both positions are wrong in the same way. They assume all messages carry the same risk, when in practice the difference between a safe automated reply and an expensive one is easy to see in advance.

So draw the line explicitly. Here is a working version of it.

Safe to automate: facts you have already written down

If the answer exists in your business profile and does not change based on who is asking, an agent should answer it immediately, at any hour, in whatever language the customer used.

• Opening hours, including holidays you have recorded.
• Location, parking, landmarks, and which branch is nearest.
• Listed prices and what is on the current menu or catalogue.
• Delivery areas and typical delivery time.
• Whether a specific item is in stock right now, if your stock data is connected.
• Order status, where the order already exists in your system.

These are the messages that make up most of the volume and almost none of the value of a human reply. Automating them is not cutting corners; it is the difference between a customer getting an answer at 10pm and a customer messaging a competitor at 10:05.

Never automate: anything involving a decision

A decision is anything where two reasonable people at your business might answer differently, or where the answer commits you to something. Those must reach a person, every time.

• Discounts, price matching and “what is your best price”. An agent that negotiates will eventually negotiate badly, and consistently.
• Complaints, refunds and anything with an unhappy tone. The first reply to an angry customer is not a place to be efficient.
• Promises about dates, capacity or bespoke work — event bookings, custom orders, bulk quotes.
• Anything medical, legal, financial or safety-related, without exception.
• Requests that do not match a known pattern. Unusual is exactly the signal that a person should read it.

The test is simple: if getting it wrong would cost you money, a customer, or an apology, it is not an automation candidate.

The grey middle: ordering

Taking an order sits between the two, and where it falls depends on your business. If you have a fixed catalogue and a repeatable flow — a restaurant menu, a retail item list — an agent can genuinely handle the whole thing: walk the customer through options, confirm items and quantities, collect the delivery address, and hand a structured order to the team. The risk is low because every element was defined by you in advance.

If your orders need quoting, measuring, or a judgement about feasibility, the agent should gather the requirement carefully and stop. Collecting the details is useful work. Committing to deliver them is not its job.

Escalation is the feature, not the fallback

Most disappointing automation fails at the handover rather than the reply. The bot gets stuck, sends “an agent will contact you shortly”, and pushes a bare notification to a team that now has to open the thread, read back through it, and work out what was being asked. That is slower than having had no bot at all.

A good escalation carries three things: the full conversation, a plain statement of what the customer is actually asking for, and a recommended reply the person can send, edit, or reject. The human stays in charge of the decision while the tedious part — reconstructing context — has already been done.

It also matters that a person can step into a live thread and have the agent stand down for that conversation until it is handed back. Customers should never experience two voices talking over each other.

Set the guardrails before you set it live

Write down what the agent may never promise. Set the hours in which it replies at all — some businesses prefer notify-only overnight, so nobody gets a cheerful automated answer at 3am about a complaint. Decide whether it may take orders. Then read every thread for the first week and correct the answers that are not quite right, because that week is where nearly all of the tuning happens.

Done this way, automation is not a replacement for your team. It is a filter that gives them back the hours currently spent typing the opening times, so the conversations that actually need a human get one faster.