AI Agent or Auto-Reply for Customer Messages: Which Does Your Business Need?

An auto-reply sends one fixed message. A rules flow, such as a menu or a keyword bot, sends a fixed answer down a fixed branch. An AI agent reads what the customer wrote, answers from your own information, records the request where you work, and passes the conversation to a person when it should. Most small businesses should start with the first, move to the second when the same ten questions keep arriving, and build an agent only when customers ask varied questions, when each lead has to land in your CRM, and when you want answers outside office hours without losing the human option. Below: what the free tools already do, a two-week count that tells you which rung you are on, what a good agent does, and the limits to know.

Start with what is already in your hands, on any channel.

• WhatsApp Business app. WhatsApp's page for the app (as captured on 15 April 2026) says it lets you “greet people the moment they start a conversation and respond to messages even when you're away”. The app is aimed at the small business owner.
• Email. Gmail's help (as captured on 8 March 2026) describes a vacation responder: when someone sends you a message, “they receive an automatic reply”. It is the same idea on the channel your quotes and invoices arrive on.
• The phone and the website. A missed-call message, a form that confirms receipt and says when you will reply, and an opening-hours line on the contact page are all auto-replies too.

For a business that gets a modest number of messages, set the greeting, the away message and a handful of saved replies first, and see how far they go. This costs nothing and it tells you what customers actually ask.

Do a two-week count before you buy anything. It takes ten minutes a day.

1. Copy every inbound message from every channel into one sheet: date, channel, the question in a few words, and what you answered.
2. Sort each into one of four groups: the same question as before, a varied question you could still answer from your own notes, a request that needs a person or a decision, and something you could not answer at all.
3. Look at the shares. As a rule of thumb, not a statistic: if most messages are the same ten questions, saved replies and a rules flow will do. If many are varied but answerable from your own documents, an agent starts to pay off. If most need a person, no tool fixes that, and the better investment is a process change.
4. Count the lost time too. Note how many arrived outside office hours, and how long customers waited. Israeli business hours run Sunday to Thursday with a short Friday, so evenings and weekends can hold a large share.

The count also gives you the 20 real questions you will need later to test any tool.

• The questions vary. A saved reply answers a question you predicted. It does not understand “is the blue one in stock in Haifa before Thursday?”
• Leads need to land somewhere. The app keeps conversations in a phone. An agent can write each enquiry into your CRM or sheet with the name, the need and the status.
• Nobody can answer at night. An away message promises a reply. An agent can give the answer.
• Several people answer. Shared numbers and several staff need rules for who answers what, and a clean handover.

1. Understands the intent, even when the customer phrases it differently from your saved replies.
2. Answers from your information only: your prices, hours, policies and stock, not from general knowledge. When it does not know, it says so.
3. Does the routine task: books a meeting, takes the order details, gives the status of an order.
4. Writes it down. The conversation and the extracted details go to your CRM or sheet.
5. Hands over to a person at the right moment, with the whole conversation, so the customer does not repeat themselves.

The handover decides whether customers like it. Design it first: which words, which situations and which amounts always go to a human. A US regulator's 2023 report on chatbots in consumer finance warns that “deficient chatbots that prevent access to live, human support can lead to law violations, diminished service, and other harms”. That is a US report about banks, not Israeli law, but the lesson travels: a customer who cannot reach a person is a customer you lose. Give the agent an exit that always works, and tell customers when they are talking to an automated assistant.

• You answer for what it says. Whatever the agent promises, the customer heard it from your business. Fence it to your own materials, and have it quote the price list rather than invent a discount.
• Platform rules differ by channel. On WhatsApp, section 4.7 of the Meta Terms for WhatsApp Business Platform (last modified 23 September 2026) bars AI providers, meaning developers of AI technologies, from using the platform to offer those technologies when they are “the primary (rather than incidental or ancillary) functionality”, and Meta decides what counts as primary. It allows AI technologies in certain countries, listed on a Meta page that sets fees for them, and Israel was not on that list when we read it on 15 October 2026. The same section says you may retain an AI provider as your Solution Provider, but may not let platform data be used to train AI models, with a limited exception for fine-tuning a model for your exclusive use. Our reading, which is not legal advice: the rule targets general-purpose AI assistants offered as the product, and a task-bound assistant that serves a company's own customers is a different thing. Meta can still decide otherwise, so keep the agent on your own business, never a general chat service, and read the terms of each channel before you build and again before you renew.
• Official routes only. On WhatsApp, unofficial libraries that imitate the app can get the business number banned.
• Customer personal data. Conversations hold names, phone numbers and sometimes health or payment details. Where they are stored, which vendors process them and how long you keep them are legal questions as well as technical ones. Ask counsel for the reading under Israeli privacy law.

Use the count from above. Take 20 real messages, write the answer you expect for each, and add some that should not be answered by the agent at all: a complaint, a request for a refund over your limit, a question outside your business. Run them, read every reply, and fix the failures. Run them again after every change. The company brain post describes a balanced version of this test, and when an automation breaks covers who gets the alert when something fails.

The agent we price today runs on WhatsApp's official Cloud API: from ₪9,500, plus an optional ₪450 a month for monitoring and monthly tuning, plus VAT as required by law. It includes a lead log with an alert on every lead, human handoff, and Hebrew and English; connecting your CRM is an optional module. Other channels are scoped on a call. Meta charges per message it delivers on the official route, as covered in how much a WhatsApp bot costs in Israel.

See the WhatsApp automation service, or book a free intro call and bring your two-week count. We will tell you whether saved replies are enough for you.

• WhatsApp Business App, WhatsApp Business, page as captured by Common Crawl on 15 April 2026. Vendor page, quoted for what the vendor says.
• Send an automatic reply when you're out of office, Gmail Help, as captured by Common Crawl on 8 March 2026.
• Chatbots in consumer finance, Consumer Financial Protection Bureau, published 6 June 2023, as captured by Common Crawl on 10 April 2026. A US report on banking, cited for the handoff lesson only.
• Meta Terms for WhatsApp Business Platform, Meta, last modified 23 September 2026, section 4.7, read on 15 October 2026. The old address, whatsapp.com/legal/business-solution-terms, now redirects here. The list of countries where AI technologies are offered, with fees, is on Meta's AI providers pricing page, read on 15 October 2026.