An AI agent instead of the admin panel: running a catalogue by voice and text

Admin panels were designed around one logic: one action, one form. Add a product, change a price, process an order. With two hundred products that works fine. With two thousand items and three suppliers using different price formats, every operation is multiplied by hundreds, and a different arithmetic begins: not «how do I do this» but «how many hours will it take».

A typical shop spends 15–30 hours a month on this routine. The most expensive part is not even the time, it is the delay: a new price list sits for three days until somebody has a free slot, and for those three days you sell at old prices or do not sell at all.

The agent is not a chatbot for customers — it is an assistant for you, inside the project. It has access to the same data as the admin panel and performs whole operations on it:

• Product parsing. Pulls names, specifications, descriptions and photos from a supplier's site and converts them into your field structure.
• Importing any price list. Excel, CSV, XML, YML or an API: the agent maps arbitrary column names to catalogue fields itself, matches items by SKU and updates prices and stock.
• Order processing. Checks availability, calculates the total and shipping, sets the status, prepares documents, writes to the customer.
• Filling filters and categories. Extracts brand, power, size and compatibility from descriptions and sets the attributes — filters stop being empty.
• Targeted bulk changes. «Raise this brand's prices by 7%», «hide everything out of stock», «move this subcategory» — one sentence instead of an hour of clicking.

Most CMSs have an import feature too — which is exactly why the question is fair. The difference is that a classic import demands a consistent format: you map the columns once, and any change in the supplier's file breaks the process. Someone renamed «Price» to «Price, EUR» and the import fails.

The agent works by meaning rather than by a rigid mapping: it understands that «Price, EUR», «Retail» and «RRP» are the same field, and that «Article», «Code» and «SKU» are the matching key. New suppliers are therefore connected without reconfiguration — and without a developer.

Voice is not about futurism, it is about the fact that an owner rarely sits at a desk. They are driving, in the warehouse, at a supplier. Saying «hide everything out of stock» from a phone takes seconds; finding the same filter in a table of thousands of rows on mobile takes minutes and patience.

One important nuance: voice only makes sense with confirmation. The right sequence is a command, then a preview of the consequences («affects 214 products, average price change +7%»), and only then execution. Without that step, voice control turns into an accident generator.

A website with a built-in AI agent starts from €1500 — catalogue, admin panel and the agent with parsing, price imports and order processing. An agent for an existing site, where database or API access exists, costs from €1200.

On top comes the model API and support — from €50 a month, depending on data volume. Work out your own case simply: 20 hours of routine a month × the hourly cost of your manager. If that exceeds a hundred euros a month, the solution pays for itself within a year even on a pessimistic estimate.

The owner's main fear is «what if it ruins the catalogue». The fear is justified, so the safety mechanics are part of the project from the start:

1. Draft before publication. Every bulk change is first shown as a list: add this many, update this many, hide this many.
2. Price deviation limit. If an import moves a price by more than the set percentage, the agent stops and asks. A broken supplier file cannot wreck the shop.
3. Log and rollback. Every action is recorded with its cause; restoring the previous state is one button.
4. Role-based rights. Products for the content manager, prices and orders for the director.
5. A human at the end. Where a mistake costs money, the agent prepares and a person confirms.

The cheapest way to test the idea is not a project but a single file. Take the worst price list among your suppliers: the one where columns are named at random and half the product names are abbreviated. If the agent reads that one correctly, it will read the rest.

We build that prototype for free: you send the file, we show what came out of it in the catalogue structure. The decision about a full project is then made on a result, not on promises.