An AI assistant is software that answers questions from customers and staff using your company's own documents. The model is not retrained on your files: for every question the system finds the relevant fragments in a knowledge base and builds the answer from them, citing the source. What sets it apart from a button-driven bot is that it understands a question asked in plain words and can admit that the documents hold no answer.
Key takeaways
The difference is not intelligence but where the answer comes from: a bot shows pre-written text behind a button, an assistant assembles an answer from your documents for the specific question. Everything else follows from that.
| Point of comparison | Button-driven bot | AI assistant |
|---|---|---|
| How a question is asked | By picking a menu item | In free text or by voice |
| Where the answer comes from | From a pre-written script | From document fragments matched to the question |
| Conversation context | Every step stands alone | Remembers earlier messages in the dialogue |
| An unfamiliar question | Runs into "no such option" | Answers by meaning or hands over to a person |
| What breaks it | An overgrown menu | Outdated documents that contradict each other |
| What a launch requires | A script and ready texts | A script, a knowledge base and answer rules |
You do not have to pick one of the two. Buttons work well where the path is known in advance — booking, payment, order status — while the assistant takes the open questions. What scripted bots can do without AI is covered in our article on Telegram bots for business.
The design is the same whoever built the model under the hood. Documents are split into fragments in advance and loaded into a search index. A question arrives, the system finds the fragments that match it by meaning, passes them to the model together with the question and the answer rules, and the model writes the answer from those fragments.
The approach is called retrieval-augmented generation. Google states the point plainly in the description of its own service: context from your private data is there so that the model hallucinates less and answers questions more accurately.
Hence the practical conclusion: for an assistant to "know" your product you need neither your own model nor a training cycle. You need documents in readable form and a rule about which one counts as current. How a knowledge base, answer rules and fine-tuning differ is covered in our article on integrating AI into a CRM, ERP and apps.
A knowledge base is not a folder of files but a set of documents, each with an owner, a date and a status. That is usually what is missing: the assistant honestly answers from whatever it found, and if the company keeps three versions of a price list, it will take any one of the three.
Contradictions are sorted out before launch, not after the first customer complaint. The working rule is simple: one topic, one current document with an owner and a review date, while earlier versions go to the archive and never reach the index.
Scans and photos are a separate job: they are recognised, checked and only then added to the index. How that works and where a person is still needed is covered in our article on document processing with AI.
Answers are not the whole story: with access to company systems the assistant performs actions and replies using a specific customer's data rather than a generic service description.
The same assistant runs in several places at once: on the website, in Telegram, in a mobile app, in a customer portal and in the company's internal chat. The knowledge base and the rules are shared; only the presentation and the set of permitted actions change.
An internal assistant differs from a customer-facing one not in technology but in the contents of its base and its rights: staff see process rules, templates and customer data, a customer sees only what concerns their own order.
The assistant must not show a person anything they have no right to see without it. Rights are configured before launch — by the employee's role and department, by document ownership and by customer type — and every request is written to a log.
Important Customer correspondence and CRM records are personal data. The Law on Personal Data No. ZRU-547 treats collection, storage and use of such information as processing and allows it on one of the grounds in Article 18, most often the person's consent. If customer questions travel to a cloud model abroad, that is a cross-border transfer, and its conditions are set out in Article 27-1 as amended by Law No. ZRU-1125 of 26 March 2026.
Biometric and genetic data, along with data on users of telecom operators' services, must be stored in Uzbekistan. Other personal data may be processed abroad if the conditions of the same article are met, and that is what the choice between a cloud model and your own server comes down to. Roles, the access log and protection of the customer base are covered in our article on access rights in a CRM.
One more decision is what the assistant does with personal data inside the dialogue itself: whether it masks numbers and names before sending a request, how long it keeps the conversation and by what rule it deletes it.
A language model always produces plausible text, and with a weak or contradictory source that plausible text turns out to be wrong. This cannot be ruled out completely; the risk is reduced by several measures.
The law sets one more boundary. Law No. ZRU-1115 of 21 January 2026 forbids relying solely on the conclusions of artificial intelligence systems when taking legally significant decisions that affect a person's rights and freedoms: a refusal, a charge or an HR decision is signed by an employee.
The work goes step by step, and the first two steps are not about technology. Until the boundaries of the answer are written down and the documents are gathered, there is nothing to configure.
The third step usually takes the longest: documents have to be collected from different people and reconciled with each other. It also decides the outcome — the assistant will never know more than your materials say.
An assistant has to close a measurable task, so the metrics are fixed before launch and compared with the same period before it.
The last item is the most useful: a list of unanswered questions is a ready plan for extending the base. After a month of work it usually explains why customers end up with a manager better than any survey.
The price depends on what exactly you are launching: a process review, a bot for taking requests or an assistant built on a knowledge base. On the service page these are separate formats with separate starting prices.
Three things push the budget above the starting figure: the size and state of the knowledge base, the number of channels and systems to connect, and the depth of actions the assistant performs itself. What each format includes is described on our bots and AI assistants page.
After launch there are running costs: payment for model requests or a server if the model is your own, plus the person who keeps the knowledge base current. Support for a company's business systems costs from $700 a month; the scope is described on our technical support page.
Syntra Systems starts with enquiries and documents rather than with the choice of model: we work out what customers ask about and what staff look for, separate what closes without a person, and only then assemble the knowledge base and the answer rules.
We configure the assistant to rely on your materials and to say honestly when it has no answer. We test it on real questions, collect errors and fix them before customers get access, and after launch we watch the dialogues and update the knowledge base so the answers do not go stale.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
What counts is topics rather than files: to start, cover the questions that repeat most often — the price list, terms of work, frequent support questions and a couple of process rules. The rest is added after launch from the list of questions the assistant could not answer.
It takes over repetitive operations: standard questions, information lookup, draft replies. Complex, contentious and high-stakes cases stay with a person, and under the law on artificial intelligence a legally significant decision cannot rest on the system’s conclusion alone.
That's checked on your real enquiries, not on claims: before launch the assistant runs against actual past questions, and the error rate is counted separately for each language. If Uzbek shows noticeably more mistakes, those topics are handed to a human at the start.
The document owner on the company side: when prices or a process change, the file is updated, not the assistant’s settings. Without a named owner the base goes stale within months and the assistant starts confidently answering on old terms.
Drive search matches words and returns a list of documents you still have to work through. An assistant searches by meaning, assembles an answer from several sources and shows where it came from, so an employee gets an answer rather than a folder of materials.
That counts as a cross-border transfer of personal data — it needs a legal basis, usually the person's consent, and it has to meet the law's conditions for sending data abroad. Biometric and genetic data can't be sent at all and may only be stored inside Uzbekistan, while names and phone numbers are best replaced with placeholders before the request is sent.
An assistant built on your own documents starts from $3,000; if it isn't yet clear which tasks to hand to AI, we start with an AI audit at $400, which delivers an automation plan. After launch you keep paying for requests to the model, or for a server if you run your own model, plus the time of whoever updates the knowledge base; ongoing support for working systems starts from $700 a month.
