Guide · AI on WhatsApp · For the owner about to switch Meta's agent on
What to give Meta's AI agent to learn from on WhatsApp: business details, FAQs, files and your website
Meta's AI agent answers customers from what the business gives it. Meta's developer pages name four sources, and say quite precisely how each should be written. Get them right and the agent quotes your policies; get them wrong and it guesses, or quotes the wrong country's prices.
Published 7 October 2026 · Updated 8 October 2026
Four places the agent learns from
- Business details
- One set per number: how customers can pay, the return policy, how to buy, delivery and shipping, a short description of the business, and the email, opening hours and address.
- FAQs
- Question and answer pairs. Meta calls them the most direct way to control how the agent answers a known question: it finds the matching entry and answers from it, instead of working an answer out from everything else.
- Files
- Longer or more varied material, such as product guides, policy documents and price lists. Meta says the agent draws on them when no FAQ matches the customer's question.
- Your website
- Meta crawls the site and adds what it finds, and crawls it again from time to time so the agent keeps up with changes.
Writing FAQs the way Meta says works
- Ask it as a customer would. Meta's example: "What is your return policy?" works better than a heading such as "Returns, exchanges, and refunds", because the agent matches entries more accurately when they sound like real questions.
- One question per entry. Each FAQ should cover a single, specific topic.
- Make every answer stand alone. Factual, short and complete, so the customer does not need to ask again. Meta says not to point to other entries, because each one is read on its own.
- Start with the questions you hear most. Meta warns that going beyond a few hundred FAQs can make it harder for the agent to find the right one.
- Change the wording with care. Rewording a question changes which customer messages it is matched to; rewording the answer changes what the agent says.
Where the questions come from
Files and the website
Files can be PDFs, Word documents or images in PNG or JPG. Spreadsheets in CSV or Excel are accepted only when that option is turned on for the number. Meta sets a limit of 100 million bytes, roughly 100 MB, per file, and rejects a file whose name ends in a different type from its contents.
A website is crawled in full by default. Meta lets the crawl be limited to certain parts of the site, certain address patterns or a list of exact pages, and says this matters most for sites with several regions or languages, where crawling everything gives the agent contradictory content, such as prices or policies from a market the customer is not in.
Our reading for UAE businesses
Checking what it picked up
Meta reports each website crawl as finished, finished with nothing usable found, or failed, with a count of the pages read. Its examples of a failure are a site that cannot be reached and a site whose settings (a file called robots.txt) block Meta's crawler. Meta suggests reviewing what the agent picked up and then widening or narrowing the crawl.
Removing a source stops the agent using it and leaves the others alone. Saving the business details replaces the whole set, so our reading is to keep one master copy and edit that. To see how the agent then answers, use the testing guide before launch and the reports guide after it.
Questions
Where do we start when writing FAQs for the agent?
With the questions you already answer over and over. A WhatsApp Business guide suggests writing a prepared answer to each and using them as quick replies; Meta says FAQs and pre-saved replies can be used to guide Meta Business Agent so it speaks in your voice.
How many FAQs should we give the agent?
As many as customers really ask, starting with the most common. Meta warns that beyond a few hundred the agent can find it harder to pick the right one.
Can the agent read our price list?
Yes, as a file. PDFs, Word documents and PNG or JPG images are accepted; CSV and Excel spreadsheets only when that option is turned on for the number. Meta says files are used when no FAQ matches the question.
Our website is in Arabic and English. Is that a problem?
It can be. Meta says crawling a whole multi-language or multi-region site can give the agent contradictory content, and lets the crawl be limited to the parts that fit.
Why would Meta fail to read our website?
Meta's examples are a site that cannot be reached and a site whose robots.txt file blocks its crawler. The crawl status says which, along with how many pages were read.
Sources
- Meta's developer documentation: Meta Business Agent Agent Knowledge, Business Info, read 7 October 2026.
- Meta's developer documentation: Meta Business Agent Agent Knowledge, FAQs, read 7 October 2026.
- Meta's developer documentation: Meta Business Agent Agent Knowledge, Files, read 7 October 2026.
- Meta's developer documentation: Meta Business Agent Agent Knowledge, Websites, read 7 October 2026.
- WhatsApp Business: Grow your business without sacrificing customer satisfaction, read 8 October 2026.
More plain-language notes like this one are on the guides page.
Want every chat in one place you own?
We set up a business's WhatsApp number on a system its team runs, so every customer conversation sits in one place the business owns. A short call is the easiest way to see how that would work for you.
Book a demo