What drives the cost of an AI shopping assistant?

What you are paying for

A useful AI shopping assistant is not a chat box glued onto a shop. It is a small system that reads your catalogue through tools, follows your rules, stays within a budget and is tested like any other feature. Its cost depends far more on your data and the actions it must take than on the AI model. The sections below explain the factors, using what we learned building the assistants of Cumpără cu AI and Raftul copiilor.

After launch, the running costs are model usage (billed by the AI provider to your account, with a daily cap), your hosting, and any maintenance you choose. Our care plans cover routine maintenance; changes beyond them are quoted separately.

1. Catalogue quality sets the ceiling

An assistant can only be as precise as the data it reads. If sizes, ages or specifications are missing from half your products, the assistant must say “unknown”, or a separate job must extract them first. On our sites, specifications are read from titles and descriptions and a missing value is stated as unknown, never as absent.

Ask early: are categories consistent, are key attributes structured, are prices and stock fresh? Cleaning or enriching data is often the largest hidden part of the budget, and it is better scoped as its own step (see AI workflow automation).

2. The tools decide what it can do

The model never queries your database directly. It calls tools you define, and each tool is code to write, secure and test. A typical first set:

• Search and filter the catalogue, with the same filters a visitor has.
• Product details, including what the shop says about compatibility or contents.
• Compare two or three products on facts both state.
• Ask a clarifying question when the request is ambiguous, instead of guessing.

3. Rules, refusals and honesty

Most of the engineering goes into what the assistant must not do: invent products or prices, promise discounts, disparage a brand, answer off-topic questions at your expense, or claim it placed an order. On our sites, plain greetings and off-topic questions are answered without a model call, prices are written by the application, and answers that break a rule are rewritten before they are shown.

4. Model usage and limits

Model fees are usually modest for a well-designed assistant, because the tools do the heavy lifting and a low-cost model handles the language. Our live assistants run on an inexpensive Gemini model with a backup. What matters is control: per-visitor limits, a site-wide daily cap with alerts, and abuse protection. Raftul copiilor and Cumpără cu AI use per-minute and daily limits, canned answers for junk input and a self-hosted proof-of-work challenge only when needed.

5. Evaluation is part of the product

You need a written set of realistic questions with the expected behaviour, run before every change to instructions, tools or model. We keep such sets for search (290 cases on Cumpără cu AI) and test the assistant end to end on a staging copy of the database, never on real visitors’ conversations.

6. What moves the budget

The smallest useful pilot is one catalogue in one language with search, details, comparison and clarifying questions, plus limits and an evaluation set. More languages, more tools, data cleaning or integration into a complex existing platform make the project larger. Model usage is paid directly to the provider.

7. Custom assistant or subscription chatbot?

A subscription chatbot is often enough when you mainly need answers about delivery, returns and opening hours, or product search on a standard shop platform the service already supports. It costs less to start and someone else maintains it.

A custom assistant earns its higher starting cost when it must work through your own catalogue logic or several data sources, follow rules a generic product cannot express, act on your page (filters, comparisons), or be tested against your own scenarios before every change. You also own the result: the code lives in your repository and the accounts are in your name. If a subscription covers your case, we will tell you so.