AI assistant for business: how it answers from your own data instead of guessing prices

An AI assistant is a language model wrapped in your data and rules

A small business owner reading about chat assistants usually asks one thing: can a program answer customers at any hour without making things up. An AI assistant for business has three parts. One is an interface: a chat panel on a site, in a customer account or in a messenger. Another is a language model that writes the replies. The third is the company's own data and rules, which decide what is true and what may be promised.

A scripted bot follows buttons and replies written in advance: it never says anything it was not given, and it gets lost when a question is phrased unexpectedly. An assistant reads free text and writes a new reply each time, which is both its strength and its risk. An agent goes further and acts: the OWASP GenAI Security Project describes LLM systems granted a degree of agency, able to call functions or other systems. This story is about the assistant that suggests.

Why a language model fills the gaps with invented prices

The 2025 OWASP Top 10 of risks for applications built on large language models lists misinformation as LLM09: false or misleading information that appears credible. A major source is hallucination, content that seems accurate but is fabricated. The cause is that models fill gaps in their training data using statistical patterns. A missing fact produces not silence but a plausible sentence.

For a business the gaps sit where mistakes are expensive: price, delivery terms, stock, the guarantee. OWASP cites Air Canada, whose chatbot gave travellers wrong information; the airline was successfully sued as a result. Model makers say the same about their own work. Google's model card for Gemma 3 states that the models are not knowledge bases and may generate incorrect or outdated factual statements.

The remedy is called grounding. Google Cloud's documentation defines it as the ability to connect model output to verifiable sources of information and says it reduces hallucinations. OWASP's advice agrees: retrieve verified information from trusted databases and keep human oversight for critical facts. So the assistant looks the answer up before it speaks, and some questions never reach the model at all.

Hosted API, own server or fixed answers: three building choices

The first choice is where the model runs. A hosted service receives each question over the internet: there is no hardware to keep, but the questions leave your systems and the provider's terms apply. A self-hosted model runs on a machine you control; Ollama, an open-source tool under the MIT licence, does this on macOS, Windows or Linux. Google describes its Gemma models as lightweight open models small enough for laptops, desktops or your own cloud. Whether a small model answers your customers well is a matter for testing, not assumption.

The third choice is where the facts come from. Retrieval finds relevant passages in your own documents and passes them to the model with the question. It can match keywords or meaning: Google's Gemini API documentation describes embeddings as numerical representations for semantic search, more context-aware than keyword-based approaches. For price, returns and guarantees, a fixed reply read from the price list beats any generated sentence.

Sign-in, message limits and prompt injection in a customer assistant

An assistant that reads text can be talked into things. OWASP puts prompt injection first, as LLM01:2025, and describes it as user prompts that alter the model's behaviour or output in unintended ways. The indirect kind arrives from external sources, such as websites or files the assistant reads. Among the OWASP mitigations are constraining the model's role and clearly marking untrusted content so that it cannot pass for instructions.

Every question also costs computing time. OWASP's LLM10:2025, unbounded consumption, covers excessive and uncontrolled use that leads to denial of service and economic losses, and lists rate limits and quotas per user among the defences. Requiring sign-in is the simplest quota, because the system knows who is asking. The entry on excessive agency adds the last rule: a person approves high-impact actions before they happen, so the assistant suggests and the customer presses the button.

The VITON ID cabinet assistant: catalogue first, model last

VITON13 Studio built such an assistant for its own site. It is the studio's own product, not a client commission: the VIT LIVE tab of the signed-in VITON ID cabinet, with a panel titled VIT ASSISTENT. Members should get real items with real prices in their own language, not a model's guess. An earlier version gave the model only site-search hits, meaning titles and paths, so when asked how much a website costs, it hedged or invented a figure.

The rebuilt assistant tries three layers in a fixed order, and the first that matches gives the answer. First comes the catalogue: 84 items for each language, 32 products from Old Money, RISE, ANIRI and VITON XIII and 52 studio packages from 13 service pages. For questions about buying, products, services, the store or the bag, a keyword matcher replies with up to six cards, each with a price, labelled VITON13 CATALOG, and no model is called. Next come fixed replies, with exact answers on prices, ordering, discounts, bonuses and reply times read from the service price ladders and the bonus rules. The self-hosted model comes last.

What the model sees in the VITON ID assistant and what it may not do

The model is served by Ollama, by default on the same server as the site; the code names gemma3:4b as the default and lets the server's environment choose another. It receives the question, the two messages before it and a fact sheet generated at that moment from the site's pricing and ordering data. Questions about the site add up to eight results from the site's keyword search. Questions about news, weather or exchange rates add up to two web results, passed as untrusted evidence, never as instructions, and listed in the panel as sources.

It must answer only in the member's language, in one to three sentences, and when the fact sheet lacks a figure, say that the price is confirmed after the brief instead of guessing. The temperature is 0.1; the Ollama documentation gives 0.8 as the default and says raising it makes answers more creative. Replies are capped at 180 tokens by default, where the Ollama default is no limit. The text streams into the panel as it is written.

Every request is checked for a signed-in session: on 19 September 2026 the live route turned away a visitor who was not signed in with 401 REGISTRATION_REQUIRED, as intended. Each member can send 8 messages in 5 minutes. On a card the member presses Choose, checks the options and only then presses Add to bag; the assistant never adds anything by itself. Eleven automated tests of the model call and its error when the server is down, the fact sheet, the web search rules and the fixed replies all passed on 19 September 2026.

What the VITON ID case cannot tell you about AI assistants

Here is where the evidence stops. The assistant is not a public chat widget: only members signed in to VITON ID can use it, and there are no usage or conversion figures to report. It does not take orders or pass requests to the studio; those go through the forms on the service pages. Nor does the case show the model running in production or judge the quality of its answers.

Matching is by keywords, not meaning: there are no embeddings and no vector store. The word lists have gaps and differ between languages, so some questions, including a few of the panel's own suggestions, miss the right cards. The catalogue holds 3 of the 5 VITON XIII fragrances and no VIT MARKET goods, and the fixed answers on prices and ordering exist only in Russian and English. What carries over to another business is the order of the steps, not a promise about the model.

Preparing questions, prices and people before an AI assistant goes live

Three things decide the result more than the model does. First, the questions customers actually ask, taken from mail, chats and call notes, with the ones that need an exact answer marked. Second, one price and terms sheet as the single source of truth, kept current by a named person. Third, who takes over when the assistant cannot help. An AI assistant for business built without these repeats old problems, only faster.

An off-the-shelf chatbot is enough when the questions are few, repeat often and rarely change: opening hours, an address, a short list of common questions. A custom assistant is worth ordering when answers depend on a live catalogue, prices must be exact in several languages, access is tied to customer accounts or the assistant reads from your own systems. VITON13 Studio does this kind of work. The linked guide explains what an AI chatbot for a website costs, and the case page shows the VITON ID assistant in detail.