Visitors fill in your contact form with questions you already have answers to. Do you ship to Jeddah? How long is the warranty? Does the service include installation? What papers do I need to register? The answers sit in a catalog, a policy or a PDF, but the visitor will not read twenty pages, and your employee replies the next day.
An assistant that answers from company documents has become one of the most requested AI projects, and it does help. It also has limits you should understand before you put it in front of your customers.
How does an AI assistant answer from your company documents?
The assistant works with a method usually called RAG. For every question, it searches your documents for the closest passages, hands them to the language model, and tells the model to answer from them alone. The model reads the right passage at the moment of the question. It is not trained on your documents in advance.
This has a big practical advantage. When you change a price or a policy, you update the document and the answer changes right away, with nothing to retrain. The assistant can also cite where the answer came from, so the visitor or employee can see the source.
- Visitor's questionIn Arabic, English or dialect
- Search the documentsThe closest passages from company files
- Write the answerFrom those passages only, with the source
- No answer found?It says so and offers to connect to a person
Which documents should the chatbot answer from?
Use the documents a good customer service employee would turn to: the list of services and published prices, shipping, returns and warranty policies, FAQs, registration terms, working hours and branches. If your system has an API, the assistant can also connect to live data such as order status or available appointments.
An answer can only be as good as the documents behind it. If you have two files with two different prices for the same service, the assistant will quote one price one time and the other price the next. Half the work in these projects goes into organizing the documents and removing old ones, and that work helps your company even if you never launch the assistant.
What are the limits of an AI website assistant?
The assistant sometimes gets things wrong, even when it is restricted to your documents. It may read an ambiguous question the wrong way, or combine two pieces of information from different places and reach an inaccurate conclusion. These are the main limits we are upfront about with clients before we start:
- It should not make decisions that commit the company, such as approving a refund or offering a discount.
- It does not know what is not written down. Experience that lives in your employee's head is out of its reach.
- It needs clear instructions to say "I don't know" instead of guessing.
- It needs follow-up after launch, because real questions expose gaps in the documents that nobody expected.
How do you stop a chatbot from making up answers?
You combine several layers. We restrict the assistant to answering from the documents, require it to cite its source, give it a clear sentence to use when it cannot find the answer, and block it from discussing topics outside the company's business. Then, before launch, we test it on at least a hundred real questions taken from your customers' past messages and read the answers one by one.
After launch, review two kinds of conversations every week: those where it said "I don't know," and those where the visitor asked for a person. The first tells you what is missing from the documents. The second tells you where the answer did not convince.
Put a short line under the chat window explaining that answers are automated and important matters may need confirmation from the team. This sets visitors' expectations and protects you if a mistake happens.
Should the assistant run on your website, WhatsApp or internally?
The same engine can run in more than one place. On the website it answers new visitors and collects details from anyone who wants a quote. On WhatsApp it serves existing customers, and WhatsApp has its own rules, which we explain in automated WhatsApp replies. Inside the company it answers employees from the procedures manual and internal policies, which carries less risk and makes a good starting point.
If many of your visitors arrive through answers from AI search tools, organizing your documents also helps with getting your company into AI answers.
What determines the running cost of an AI chatbot?
The monthly running cost depends mostly on how many conversations there are and how long they run, because model providers charge by the amount of text processed. Hosting and any connections to other systems come on top. An assistant on a service company's website handling dozens of conversations a day costs very differently from one on a large online store handling thousands. Estimate your conversation volume from the messages you get now, and ask the vendor for an estimate at that number and at double it.
What should you prepare before building a chatbot?
- Approved documents only, with one version per topic.
- A list of the most frequent questions from email, WhatsApp and website forms.
- A clear decision on when the assistant hands over to an employee, and to whom.
- One person responsible for updating the documents when prices or policies change.
We build these assistants on company documents and connect them to websites, WhatsApp and internal systems as part of our AI and automation solutions service. Send us your website link and the questions you get most, and we will tell you what the assistant could answer and what stays with your team.
Questions our partners ask
Does the assistant need a model trained on my company data?
In most cases, no. The common approach is for the assistant to search your documents for each question and hand the model the relevant passages to answer from. That costs less than training, and updating a document is enough to change the answer.
Can the assistant make up a wrong answer?
It can. That happens far less when it is restricted to answering from your documents and told to say it does not know when it cannot find the answer. Testing it on real questions before launch is still essential.
Does the assistant work on WhatsApp too?
Yes. The same engine can be connected to your website and to WhatsApp through the WhatsApp Business Platform, within WhatsApp's messaging rules.
Written by the Visual Innovate team, the people who build, host and follow up systems, websites and apps for our clients in Jordan, Saudi Arabia, Oman and Egypt.
