Blog / WhatsApp

A WhatsApp Auto Reply That Reads Your Docs Instead of Guessing

21 October 2026 · 7 min read · By Runix B

A WhatsApp Auto Reply That Reads Your Docs Instead of Guessing

The built-in WhatsApp auto reply fits in a sentence: "Thanks for your message, we'll get back to you." It is polite, it is instant, and it answers nothing. The customer who asked whether the oak table ships to Edinburgh still has to wait until morning.

So people upgrade to an AI reply. And that is where the real trouble starts, because a bot that answers fast and answers wrong is worse than the polite brush-off. The brush-off at least doesn't lie.

This is part two of our series on WhatsApp automation. Part one covered the overview and setup. This one is narrower: how to make a WhatsApp auto reply that answers from your own documents, says so when it can't, and passes the chat to a person at the right moment.

Written by the ChatterMate team. We build an open-source AI support agent, so weigh our product claims accordingly. We've kept them to what we could verify.

Three kinds of WhatsApp auto reply

They get lumped together. They shouldn't.

The first is the away message. Fixed text, sent when you're closed. WhatsApp Business has this built in. Free, dull, safe.

The second is the keyword menu. "Reply 1 for opening hours, 2 for delivery." It works until a customer types a sentence, which is what customers do.

The third is the grounded AI reply: the customer writes a normal question, the system finds the relevant passage in your own content, and writes an answer from that passage alone. That's the one this post is about. It is also the only one of the three that can be confidently wrong, which is why most of what follows is about guardrails.

What "grounded" has to mean on a phone

We've written at length about why chatbots invent things. The short form: a language model fills gaps with plausible text. Grounding means the model is handed your passages and told to answer from them, and to say "I don't know" when they don't cover the question.

On a website you can show a source card under the answer. WhatsApp is plainer. A message is text, a few attachments and links. Meta's cloud API caps a text body at 4,096 characters (one BSP's docs recommend staying under 1,600 for clean rendering), and formatting is limited to asterisks for bold and underscore pairs for italics, per 360dialog's message documentation. No tables. No expandable citations.

Here is something we have to be straight about. Our public README and channels page say the agent answers from your knowledge base on every channel, WhatsApp included. Neither says that a customer-facing WhatsApp reply carries a visible source citation. We haven't verified that it does, so we won't claim it. If you need a link at the end of each answer, test for it on your own number before you promise it to anyone. And if it isn't there, the fix is cheap: write the canonical URL into the source page itself ("Full delivery terms: example.com/delivery") so the answer can point at it.

That is a real design constraint, not a footnote. On a phone, a short answer with a link beats a long answer without one.

Write your docs for a thumb, not a desk

Most knowledge bases are written for people reading at a laptop. Thirty-word sentences, three caveats, a table of regions. A model will faithfully compress that into a WhatsApp message and the customer will squint at it.

What works better is boring. One question per page. The answer in the first two sentences. The exception directly after, not in a footnote. We covered the longer method in writing help-center content an AI can use, but three rules do most of the work for WhatsApp:

  • Put numbers and units in the same sentence as the thing they describe: "Standard delivery takes 5 to 7 working days," not "see the table."
  • Name the exceptions: "Orders to the Highlands add 2 days."
  • Say what you don't do. A page that states "we do not offer custom sizes under 40 cm" gives the agent something to say instead of something to invent.

That last one surprises people. Negative statements are the most useful lines in a knowledge base, and the ones teams skip.

Set up the WhatsApp auto reply, in order

The wiring is covered in part one. In short: a Meta developer app, a business number, your app secret, and the channel connected in the ChatterMate dashboard (see issue #353 if manual credentials ever send but don't receive; the fix was a per-account webhook signed with your own app secret). Skip it here. The part that decides quality is what you do before connecting.

1. Load the sources that customers actually ask about

Not everything. Start with pricing, delivery or lead times, returns, opening hours and the five product pages that get most of the questions. The full walkthrough is in how to train an AI chatbot on your docs. A small, clean knowledge base beats a big stale one every time.

2. Pull your real questions from your real chats

Open your WhatsApp history. Copy out thirty genuine customer messages, typos included. "do u deliver 2 leeds". "how long for the walnut one". These are your test set, and they look nothing like the tidy questions you'd write yourself.

3. Run the test before the number is live

Ask all thirty. Mark each answer as correct, wrong, or "didn't know." Wrong is the number you care about. A "didn't know" is a gap in your content. A wrong answer is a gap in your guardrails, and you fix those first. AI chatbot testing has the longer checklist.

4. Decide what the agent must never answer

Refunds above a limit. Anything legal. Anything medical. Complaints with the word "solicitor" in them. Write these as handoff triggers, not as things the model is asked to be careful about. "Be careful" is a hope. A trigger is a rule.

5. Connect, then watch the first fifty conversations

Read them. All of them. Fifty is a small number and it will teach you more than any dashboard.

The 24-hour window changes what "auto reply" can do

WhatsApp gives you 24 hours from a customer's last message to send free-form replies. After that you're limited to approved templates. Meta's documentation is explicit about it, and ChatterMate's README lists 24-hour session handling for WhatsApp.

For a grounded reply that's mostly good news. A customer asks, the agent answers, the window is open. The trouble is the handoff. If a human takes the chat at 17:30 and the customer goes quiet, you have until the same time tomorrow to follow up freely. After that it's a template. So when the agent hands off, it should say so plainly and set an expectation: "A person will reply here by 10am tomorrow." A promise you can keep is worth more than a clever answer.

Is an AI reply even allowed on WhatsApp?

Fair question, and it comes up. Meta tightened its WhatsApp Business rules in late 2025, effective for existing users on 15 January 2026. The target was general-purpose assistants, the ChatGPT-on-WhatsApp kind. respond.io's summary notes that business bots for support, FAQs, bookings and order tracking remain allowed, as long as the bot is secondary to a real business service. It also recommends a handoff to a human when the bot can't resolve something.

We're relying on a third-party summary there, not Meta's own wording, so read Meta's current terms before you build anything you'd stake a business number on. But a docs-grounded support agent that hands off to a person is plainly the shape of thing the rule leaves alone. A bot that chats about anything is the shape it doesn't.

Where this goes wrong

We've watched a few patterns repeat.

The stale page. The agent quotes last year's price because nobody updated the page. Grounding makes the bot faithful to your docs, including the wrong ones. Put an owner and a review date on every source.

The over-helpful guess. The customer asks about a product you don't sell and the bot improvises an answer from a similar one. Fix it with a page that lists what you don't do, and a handoff for anything unmatched.

The wall of text. A technically correct answer, nine lines long, in a chat bubble. Keep replies short and give the link for detail.

The silent handoff. The bot stops replying and no human arrives. This one loses customers fastest. Business-hours awareness helps, and ChatterMate's README does list it for handoff, but you still need to staff the inbox or say honestly when nobody's there. We went into the mechanics in chatbot human handoff that doesn't frustrate customers.

When not to use a grounded AI reply

If your questions are mostly "where is my order," the answer lives in your order system, not your documents. A docs-grounded agent can't look up a parcel it has no access to. Use a tracking link in a template, or wire the order lookup separately. Likewise if nearly every message needs a judgment call from a person, automate the collection of details and skip the answering.

Measuring whether it works

Three numbers, over the first two weeks. How many conversations ended without a person. How many were handed off, and for what reason. And the list of questions the agent couldn't answer, which doubles as your content backlog. Count resolution strictly: a customer who gave up isn't a resolved chat. Cost per resolution shows how easily that number gets flattered.

Next in the series: lead capture and qualification on WhatsApp, and what changes when the conversation starts with a stranger's enquiry rather than a customer's question.

ChatterMate is open source, self-hostable and free to start: the Free plan includes 50 AI replies a month. Try it at chattermate.chat.

Where this fits: WhatsApp

See how ChatterMate handles this in the product.

Read about WhatsApp

Plug it in

Add your website and start teaching your AI agent.

Start free