AI customer-support automation: what it can and cannot do

An AI support assistant can answer the questions whose answers are already written down, in seconds, at any hour, on WhatsApp, email or a help centre: opening hours, prices, policies, order status, how-to steps. It cannot decide a refund, resolve a complaint, or promise something your policy doesn't cover; those go to a person, ideally with the assistant having collected the details first. Set it up to answer only from your own documents, show the source, and hand over clearly when unsure. Measure the share answered without a person and the time to first reply.

What it does well

• Repeated factual questions. "Are you open Friday?" "Do you deliver to Haifa?" "What's the price for X?" If the answer is in a document, the assistant finds it and replies with the source.
• Status lookups. Order, booking or ticket status pulled from your system by a reference number.
• Collecting details before a person steps in. Name, order number, the problem in one line, a photo. The person then starts with everything they need.
• After-hours first replies. "We've got your message, here's what we can tell you now, a person will follow up at 9:00."
• Routing. Sending each message to the right person or queue, with a short summary.

What it should not do

• Decide refunds, discounts or exceptions. It can draft the reply; a person approves it.
• Handle complaints end to end. Upset customers want a person; the assistant should recognise that and hand over quickly.
• Answer from general knowledge. If your documents don't say it, the answer is "let me get a person"; an assistant that guesses about your business is the one that invents a policy.
• Pretend to be a person. Say it is an assistant; customers take it better and the law in several markets expects it.

Setting it up

1. Collect the answers. Last month's conversations, sorted by type. Write the top thirty answers in plain language; this is the assistant's only source of truth.
2. Choose the channel. Most Israeli small businesses start on WhatsApp Business Platform, then add email or the website chat. Each has its own rules; WhatsApp needs approved templates for messages you initiate.
3. Define the hand-over. Which cases go to a person, how (a WhatsApp group, a CRM ticket, an email), and what the customer is told meanwhile.
4. Test on real conversations. Forty past conversations, including the angry ones; judge each answer with the person who normally handles them.
5. Go live with a person watching for two weeks. Read every hand-over and every answer the first week; adjust the documents, not the prompt, when something is wrong.

What to measure

Report them monthly, against the month before the assistant.

Costs to expect

A build in the low-to-mid range of the cost guide, plus monthly costs for the messaging channel, the AI usage and a support plan. The written answers are the part that doesn't get cheaper with tools, and the part that pays off most.

Best fit and not a good fit

Best fit: businesses with repeated questions and written answers, or willing to write them: shops, clinics, service providers, delivery, courses. Not a good fit: businesses where every conversation is a negotiation or a diagnosis; there, use the assistant to collect details and route, and leave the answering to people.

What to do next

Sort last month's support conversations by type and count how many are answerable from written information. If it is more than a third, find a builder on the marketplace who shows a support assistant with a visible source and a hand-over path, and ask for a two-week pilot.