RomânăContact
Menu
Services

Help shoppers find the right products with an AI assistant

We build shopping assistants that turn a visitor’s needs into catalogue searches, useful filters and product comparisons. The assistant reads product facts from your catalogue through application tools, so its answers rest on your real products, prices and shops, with checks that catch answers breaking your rules.

Plan your AI assistantExplore our work

When this service fits

  • Your catalogue is large and visitors leave because they cannot narrow it down.
  • Customers ask the same pre-sale questions: which model, what size, what fits my budget.
  • Products have specifications that are hard to compare without expert help.
  • You want an assistant, but not one that makes up answers or promises discounts you do not offer.

What you receive

  • A tool layer the model must use: search, filter, product details, comparison, clarifying questions.
  • Synchronisation with the page: what the assistant filters is exactly what the visitor could filter by hand, shown as the same chips.
  • Rules for the conversation: tone, refusals of off-topic requests, missing data stated as unknown, no negative claims about brands.
  • Usage limits and a daily cost cap, so a busy day or an abusive visitor cannot run up the bill.
  • An evaluation set of real-style questions, run before each change to the prompt or the model.
  • An admin view of conversations, tool calls and costs.

Scoped separately

  • Cleaning or enriching a catalogue whose data is incomplete (see AI workflow automation).
  • Checkout, orders and customer accounts.
  • Model usage fees, billed by the AI provider to your account.
  • Additional languages beyond the first.
Example from our own products

Two live assistants that work through the catalogue

On Cumpără cu AI, a general request such as “a laptop under 3,000 lei” narrows the catalogue and the assistant asks one or two useful questions. A specific request returns a few picks, each with a reason taken from the product data. A measurable requirement like “16 GB RAM” becomes a spec filter, and the assistant says that only products stating the spec are shown.

The design principle is simple: the model does language, the code does facts. Greetings and off-topic questions are answered without a model call. Prices are written by the application, not by the model. On Raftul copiilor, the same engine was adapted to children’s products and checks the age a shop declares instead of assuming it.

You can watch both in the assistant demonstration.

Two laptops compared side by side: what changes, what the extra price buys, and grouped specification rows.
A comparison the assistant can open in place of the catalogue: what changes, what the extra price buys, every value with its source.

Cumpără cu AI · Raftul copiilor

How the work runs

  1. Catalogue review

    We check which fields exist, how complete they are and what a shopper needs to decide.

  2. Conversation design

    Typical questions, the tools that answer them, what the assistant must refuse, and what it says when data is missing.

  3. Pilot on your data

    One catalogue, one language, the agreed actions, on a staging copy of your data.

  4. Evaluation

    A written set of scenarios with expected behaviour, run against the pilot, with failures fixed and re-run.

  5. Launch with limits

    Rate limits, a cost cap, monitoring and a short report after the first weeks of real use.

What affects cost and schedule

  • Catalogue size and how consistent the product data is.
  • The number of tools the assistant needs (search only, or also comparison, availability, store rules).
  • Languages and conversation styles.
  • Expected traffic, which sets model choice and cost limits.
  • Where the assistant lives: your existing site, a new page, or a separate app.

Questions buyers ask

Should I build a custom assistant or subscribe to a chatbot service?

A subscription chatbot is often enough for delivery, returns and opening-hours questions, or for product search on a standard shop platform it already supports. A custom assistant is worth its starting price when it must work through your own catalogue logic or data sources, follow rules a generic product cannot express, act on your page, or be tested against your own scenarios before each change. The code lives in your repository and the accounts are in your name. If a subscription covers your case, we will tell you. The cost guide explains the trade-off.

Can the assistant make up a product or a price?

It can only name products its tools returned, and prices come from your data. No system that uses a language model is perfect, so we also add checks that rewrite or block answers breaking the rules, and an evaluation set to catch regressions.

Which AI model do you use?

The one that fits the task and the budget. Our live assistants run on a low-cost Gemini model with a backup in the configured order; we have also built adapters for Anthropic- and OpenAI-compatible interfaces. The tool layer stays the same when the model changes.

Does it work on phones?

Yes. On our sites the assistant is a column beside the catalogue on large screens and a separate tab on phones, and both views stay in sync.

Related services

Plan your AI assistant

Tell us what you need in a few sentences. NeedAI Studio answers personally, in English or Romanian.

Discuss your project