Advanced assistantWhatsApp AI chatbotAI customer serviceGuide

A WhatsApp AI agent that answers from your own files — built from parts, not from one long prompt

Most WhatsApp chatbots are one long page of instructions that grows until it contradicts itself and guesses your prices. We rebuilt ours from four parts — a persona, skills, your files and pictures — tried it on our own number first, and opened it to every account. How it works, what it will not do on purpose, and what our own assistant did in its first twelve days.

AAAhmed AliFounder, ExceliaBOT ··15 min read
A WhatsApp AI agent that answers from your own files, built from a persona, skills, knowledge and pictures

Every business that puts an AI assistant on WhatsApp starts the same way: a box, and a page of instructions. You are the assistant of… We sell… Our prices are… If the customer asks about delivery… It works on the first day. By the third month the page is four screens long, the old prices are still in it next to the new ones, two paragraphs disagree about the return policy, and nobody on the team dares to touch it.

Whether you call it a WhatsApp AI chatbot, a WhatsApp AI agent or AI customer service, that is the problem we set out to solve this month. The result is the advanced assistant in ExceliaBOT: instead of one block of text, the assistant is built from parts you fill in, each in its own place — and it answers from your own files instead of from a summary someone typed once. We ran it on our own WhatsApp number for twelve days before writing this, and it is open to every account.

In short

  • A WhatsApp AI agent should answer from your files — the price list, the policies, the catalogue — not from a paragraph that was true in August.
  • Split the job: who it is (persona), what it does in each situation (skills), what it knows (files) and what it shows (pictures). Each part changes without touching the others.
  • Some things are never the AI's job: working out an instalment, saying an order has shipped, promising stock. The system does those and the assistant only says them.
  • Everything is a draft you test in a chat first. Your customers see nothing until you press Save and switch on.
  • On our own number: 249 conversations in twelve days, 14 people booked a call with us through the chat, and 2 complaints went straight to the team.

Auto reply, chatbot or AI agent: what is the difference on WhatsApp

Three things go by the same name, and only one of them answers a real question:

What it doesWhere it stops
WhatsApp Business auto replySends one fixed message — a greeting, an away message — whatever the customer wrote.It cannot answer anything.
A menu chatbotOffers numbered choices: press 1 for prices, 2 for branches.The customer's question is rarely one of the choices.
An AI agentReads the message, finds the answer in your information, replies in the customer's language, and takes the booking or hands over.Only where your information stops — which is why it should answer from your files.

People search for "ChatGPT for WhatsApp Business" because they want the third kind. You do not need to connect ChatGPT or buy an API key for that: the AI is built into ExceliaBOT, and what makes it useful is not the model but what it is allowed to read.

Why one long prompt breaks

A prompt is a good way to start and a bad way to run a business. Four things go wrong, in this order:

  1. It only grows. Every new question a customer asks becomes a new line. Nothing is ever removed, because nobody remembers why it was added.
  2. It contradicts itself. The Ramadan offer is still there in Shawwal. Shipping is "two days" in one paragraph and "two to four" in another. The model picks one, and not always the same one.
  3. It cannot hold your real information. A price list of 300 items, a returns policy, the warranty terms — nobody pastes those into a prompt, so the assistant answers from a summary, and a summary is where the wrong price comes from.
  4. It cannot be tested safely. A change goes live on the next customer message. The first person to find out that it broke something is a customer.

None of this is the model's fault. It is doing exactly what the page says. The fix is to stop asking one page to do four jobs.

Four parts instead of one page

The advanced assistant: the state at the top, the four tabs, the Save and switch on button and the test chat beside them
The advanced assistant. Everything on it is a draft until you save it.

PartWhat it holdsHow often it changes
PersonaThe assistant's name, the language and dialect it speaks, its style, the welcome message and its buttons, the numbers of your departments.Rarely.
SkillsWhat it does in each situation, written as when the customer… → the assistant…: answer from a folder, take a booking, hand over to the team, tell the opening hours, work out an instalment.When your process changes.
KnowledgeYour own files in folders: price lists, policies, catalogues, FAQs, a page of your website.Every time a price changes.
PicturesA menu, a price sheet, a map, with a line saying when to send each one.Now and then.

The point of the split is that each part changes alone. A new price list is a new version of one file — the persona and the skills do not move. A new way of taking bookings is one skill — the files do not move. Nobody edits a four-screen paragraph again.

The first time you open the page it asks what kind of business you run and looks at what your account already has — a product catalogue, Store orders, a menu, the Goal page, the website chat — and builds a first version from that in a few seconds, as a draft. You start from something, not from an empty box.

Skills: one situation each

A skill is two lines in your own words. When the customer asks for a price or a plan → the assistant answers from the Pricing folder and offers a call with an expert. There is no prompt to write; the type of skill decides what the system does around those two lines.

The skill gallery: your products, order status, menu, opening hours, answer from files, collect details, offers, hand over to the team, instalment calculator, ask your system, tell your system
Pick a type, then write the two lines.

Three kinds of skill do most of the work for most businesses:

A Collect details skill: when the customer, the assistant, the fields it gathers, and what happens on the chat once the entry is saved
One skill, four boxes. The field list is what lands on the Goal page.

When two skills could fit a message, the one higher in the list wins; you move them with the arrows. A switch turns a skill off without deleting it — useful for an offer that runs only on weekends.

Train the AI on your own data: answering from your files (PDF, Word, Excel)

When a customer writes, the system looks through your files for the passages closest to the question and hands those passages to the assistant together with the message. The assistant answers from them. It does not read all 300 items of the price list every time; it reads the five lines that matter. Engineers call this retrieval (RAG); for you it means there is nothing to train and nothing to retype — you upload the PDF and the assistant answers from it.

The Knowledge tab: the Answer only from my files switch, the folders, the drop zone, and each file's buttons to read, download, upload a new version and see its versions
Folders, files, and the text the assistant actually reads.

Two lessons from building it:

A weak match is still worth showing. In our early tests we asked whether we accept Vodafone Cash. The answer was in our pricing file, but the passage scored just under the bar we had set, so the assistant was shown nothing and handed the chat over. Now the closest passages always ride along, marked as weak, and the assistant decides whether they answer the question. The same question is answered from the file today.

Read what the assistant reads. Every file has an eye button that shows the text extracted from it. A scanned PDF is a photograph of a page — it has no text, and the page refuses it and says so. An Arabic PDF can come out with words in the wrong order; we fixed the cases we found, and the eye button is how you check yours.

Then there is the switch at the top: Answer only from my files. With it on, a question your files do not answer gets your own sentence, word for word — "I'll check this with the team and get back to you" — in Arabic or English by the customer's language, and the question lands on the Goal page under Unanswered questions. That tab is your to-do list: answer the customer, add the answer to a file, and the next customer gets it from the assistant.

What the assistant will never work out by itself

Some answers are too expensive to get wrong, so the AI does not produce them. It asks the system and says what the system returned.

A draft, a test chat, then your customers

Everything you change is a draft. Your customers keep getting the replies they get today until you press Save and switch on, and the AI Prompt page keeps your old instructions — one button brings them back.

The test chat: questions written as a customer, the assistant's answers, and under each one where it came from
Under each answer: the file it read, the figures the calculator worked out, or the entry the live bot would save.

Beside the tabs there is a chat. Write as a customer — the price, the address, can I pay in instalments, and one question that has nothing to do with your business — and read what your draft answers. Under each answer it says where the answer came from: the file, the calculator, the lead it would save, or that it would hand the chat over. Fix the file or the skill when an answer is wrong and ask again. The test messages do not count against your AI replies.

We sent 319 test messages to our own draft before and after switching it on. Most of them were the same twenty questions, asked again after each change. That is what testing an assistant looks like: not one clever question, but the boring ones, repeatedly.

What our own assistant did in its first twelve days

Our own WhatsApp, Messenger and website chat have run on the advanced assistant since the evening of 24 October. It has 23 files in 3 folders — general, sales and customer service — and 10 skills: pricing, features, technical support, a free call with an expert, a sales callback, a payment or subscription problem, a complaint, general questions, wanting to subscribe, and explaining the Meta ads report.

Our own assistant, 24 October to 5 November 2026: 249 conversations, 636 replies, 14 people booked a call, 2 complaints handed to the team
Our own account, our own customers. Test threads excluded.

These are small numbers from one business, ours. They say the parts work together on real conversations; they do not say how your customers will behave.

How to build a WhatsApp AI chatbot without code, in fifteen minutes

  1. Open Advanced assistant in the sidebar, pick your kind of business and press Set up my assistant.
  2. On Persona, set the name, the dialect and the welcome line. Add the numbers of the departments it may hand customers to.
  3. On Knowledge, make a folder per subject and drop your files in — PDF, Word, Excel, CSV or text, or the link of a page on your website. Press the eye on each file once.
  4. On Skills, keep the ones the page suggested, delete what does not fit, and write the two lines of each in your own words. Add Collect details for whatever you want on the Goal page.
  5. Ask the test chat the twenty questions your customers really ask. Fix, ask again.
  6. Press Save and switch on. The next customer message on every channel is answered the new way.

The ✦ assistant in the corner of your dashboard knows this page too: ask it to add a skill for bookings, or attach a PDF and ask it to file it under Prices. Every change it proposes waits as a card with Confirm and Cancel and lands in the draft; switching it on for your customers is a card of its own. The full walk-through with every screen is in the help center: The advanced assistant: persona, skills and your files.

Frequently asked

Is there a free WhatsApp AI chatbot?

Yes. The ExceliaBOT free plan includes the advanced assistant, the files and the test chat, with no card and no time limit; it only has a smaller monthly number of AI replies.

How much does a WhatsApp AI chatbot cost?

On ExceliaBOT the AI replies are part of your plan. Meta adds its own price on WhatsApp: since October 2026 each number gets 1,000 free service replies a month, and past that Meta bills its utility rate on your Meta account (about $0.0036 a reply in Egypt). Messenger, Instagram and the website chat cost nothing extra.

Does it reply in Arabic, and in Egyptian dialect?

Yes. On the Persona tab you choose Egyptian Arabic, light Modern Standard Arabic, English, or the customer's own language, and the files can be in Arabic, English or both.

Do I need to connect ChatGPT to WhatsApp?

No. The AI is built in; there is no API key to buy and no code to write. You connect your WhatsApp number once from the Connections page.

Does it cost more than the AI Prompt page?

No. A reply is an AI reply either way and counts against the same monthly number. It is included in every plan, the free one too.

Can it make up a price?

It is told to answer only from your files and never to work out an amount by itself; instalments come from the calculator skill. Switch on Answer only from my files for the strictest setting.

How do I update a price list?

Upload the new file with the same name, or use the circular arrow on the old one. The live assistant keeps the old version until you press Save; Versions keeps the last ten, and Put back restores any of them as a new draft.

Which files can I upload?

PDF, Word, Excel, CSV, plain text and JSON, up to 10 MB each, or a page of your website by its link. A ZIP of files at once works too. A scanned PDF has no text to read and is refused.

Does it work on Messenger, Instagram and the website as well as WhatsApp?

Yes. It is the same assistant on every channel you connect, with the same files and skills.

Can I go back to my old instructions?

At any time: Stop the advanced assistant at the top of the page. Nothing is deleted on either side.

Who can change it?

The account's owners. Team members do not see the page.


Sources: the numbers are from our own ExceliaBOT account — the Goal page, the conversations and the platform's logs — from 24 October 2026, 21:03 Cairo, to 5 November 2026, with our test threads excluded. The step-by-step guide is the help center article The advanced assistant.

AA
Ahmed AliFounder, ExceliaBOT

Founder of Excelia Marketing, a performance-marketing agency working with 50+ active brands in Egypt and the Gulf, and of ExceliaBOT, the WhatsApp assistant and campaigns platform his own clients run on.

Try it on your own number, free

An AI assistant that answers your customers on WhatsApp, a shared inbox for your team, and campaigns with approved templates. The free plan needs no card, and setup takes a day.

Create your free account
← Newer post: Automate your store's offers on WhatsApp — five messages that run themselvesOlder post: Easy Orders + WhatsApp: let customers confirm their own cash-on-delivery orders →