How Does an AI Assistant Actually Work? From Canned Replies to Understanding

How Does an AI Assistant Actually Work? From Canned Replies to Understanding

Part one showed that the problem is not the company's intent but a gap in time: the customer asks, the reply comes late, and the deal goes to whoever answered first. We said the new answer is called an AI assistant. But that phrase is now sold on everything, from a miserable auto-reply to a genuine system. To decide with your eyes open, you need to know what actually happens between the arrival of your customer's message and the departure of the reply.

No jargon that does not serve you. As a business owner, three questions matter: how does it understand, where does it know your company from, and why does it sometimes get things wrong? Answering those turns you into an informed buyer instead of a buyer of promises.

Three generations, and the difference is of kind, not degree

First generation: the canned reply. One message sent to everyone who writes to you: thank you for reaching out, we will reply shortly. It understands nothing and answers nothing; its only value is telling the customer the message arrived. Most customers read it as a notification, not a reply.

Second generation: the keyword and menu bot. Type 1 for prices, 2 for appointments. It works as long as the customer walks the path you drew, and it collapses at the first sentence outside the menu. This is the generation that gave people their poor impression of automated replies: a customer types a clear question and gets back a list that does not contain it.

Third generation: the assistant built on a language model. It does not hunt for a keyword or follow a path. It reads the sentence the way an employee reads it: it understands that a customer asking how much do you take for this thing is asking about price, and that maybe tomorrow at 5 is a question about an appointment, not a confirmation of one. The gap between the second and third generation is not that the third is a bit smarter — it is that it works outside the script you wrote, which is exactly what real customers always do.

How it understands a message written in dialect, shorthand and typos

A language model is a system trained on vast amounts of text until it can judge what a sequence of words means and what should sensibly come next. The practical result is that it handles language as it actually arrives, not as we wish it arrived: in dialect, with missing words, with typos, and split across two consecutive messages.

This point matters especially in the Saudi market. Your customer does not write formal Arabic; they write the way they speak. The difference between an assistant that understands a colloquial question about visiting your branch and one that answers it with an apology for not understanding is the difference between service and embarrassment. Arabic quality is not a detail when comparing assistants — it is the first criterion.

A second point matters as much: the model holds the conversation's context. A customer who asked about a service and then, three messages later, writes and how much does it cost means that same service. A weak assistant loses the thread and asks which service do you mean; a strong one connects the beginning of the conversation to its end. That gap shows up specifically in long conversations, and it is one reason more than one tier of assistant exists.

Where does the answer come from? This is what separates your assistant from a generic one

A language model on its own knows language and the world, and knows nothing about you: not your prices, not your working hours, not your return policy. Left to answer from general knowledge, it will produce something plausible in shape and wrong in content. That is why a serious assistant works in a completely different way:

• The customer's message arrives.
• The system first searches your company's own knowledge: your services, your approved prices, your files, your frequently asked questions with the answers you wrote.
• It hands the model those passages and asks it to compose a reply grounded in them alone.
• The reply goes out in your company's tone and in the language you specified.

This mechanism is the heart of the whole matter: the assistant does not think on your behalf, it articulates what your company already knows. Which means the quality of the replies you get is a direct reflection of the quality of the knowledge you feed it — the subject of part three of this series.

Why it sometimes gets things wrong, and what prevents it

The famous failure is called hallucination: the model states, with complete confidence, something that is not in your knowledge base. The cause is simple — it is built to complete text convincingly, not to say I do not know. When it cannot find the answer in your company's knowledge, it tends to fill the gap.

So the most important question to ask a vendor is not which model is smartest, but what stops it from inventing? A serious answer comes as controls, not promises: a lock that prevents quoting any unapproved price, a list of topics it refuses to discuss, a rule that hands the conversation to a human when it does not know, and a draft mode where it writes the reply and waits for your employee's approval before sending. That deserves a full treatment, and it is part four of this series.

Why is there a fast assistant and a deeper one?

Models are not one grade. Some are fast and economical and excellent at repetitive, direct questions: working hours, published price, location, order status. Others are stronger and more expensive to run, and pull ahead when the conversation gets complicated: a customer comparing two options, objecting to a price, or describing their situation across three separate messages before asking one compound question.

The practical rule for choosing is commercial, not technical: look at your real conversations from the past month. If most of them are similar, bounded questions, the fast assistant is enough and will save you money. But if replying in your business is selling rather than answering — high deal value, a customer who needs explanation, comparison and persuasion — the deeper assistant pays its difference back on the first deal it rescues.

The Origami view

That split is exactly what we built Mahir on, Origami's AI assistant for WhatsApp. It has two explicit tiers instead of one that sells you what you do not need: Malik, the fast economical assistant for repetitive, direct questions, and Mahir, the expert assistant that holds the context of a long conversation and handles compound requests and customer objections without losing your company's voice. You move between them at any time, with no re-linking and no loss of what the assistant has learned.

Neither answers from general knowledge. Both answer from your company's knowledge as you enter it: services, prices, files, and frequently asked questions with answers you approve. And both start by default in draft mode: the assistant writes the reply and waits for your employee's approval, so you see its quality on real conversations before it says anything on its own. The trial is a full week on your own company number, with no credit card.

Conclusion

An AI assistant is neither a magic box nor a marketing trick. It is a system that understands your customer's language as they actually write it, answers from your company's knowledge rather than from the internet, and stays inside limits you set. Every remaining part of this series follows from that sentence: if it answers from your company's knowledge, the next question is how you build that knowledge in the first place, what belongs in it and what corrupts it. That is part three.

Sources

• Saudi Data and AI Authority (SDAIA) — principles and guidance on generative AI ethics.
• Official WhatsApp Business documentation from Meta — messaging and business account policies.
• Mahir official plans page — mnm.origami.sa, for tier details and the free trial.