← Back to blog

Where AI actually helps an online shop, and where it is an expensive toy

Published 4 August 2026 · 4 min read

In an online shop, AI currently pays off in three things: product copy and translation into further languages, answering repeated customer questions, and searching your own data instead of clicking through filters. It does not pay off where mistakes are expensive and checking is costly — automatic repricing, replies to complaints, and decisions about cancellations. The deciding factor is whether you can measure the cost of a single use; without that, AI gets billed as a flat fee and nobody knows whether it paid for itself.

Product copy and translation

The clearest case. A shop with a thousand products across four markets needs four thousand descriptions. By hand that is months of work; through a translation agency it costs thousands of euros — and these are texts that will change anyway.

One thing is essential: the text must be checked by a person for products where accuracy matters — composition, dimensions, safety warnings. A model can write a fluent sentence about something that is not true, and in a supplement description that is a problem, not a typo.

Answering repeated questions

Eighty percent of shop enquiries are the same ten questions: when will it arrive, how do I return it, what size is it, is it in stock. A chatbot built over your own shop data can answer them instantly, at any hour.

The boundary is clear: a chatbot should answer questions about products and process, not handle complaints. A customer writing because they are unhappy needs a person — and a chatbot answering with a template doubles the unhappiness.

Searching your own data

Less visible, but the most useful in daily work. Instead of clicking through filters you ask: "how many of these did we sell last month on the Czech market" or "which orders have been awaiting payment for more than a week".

The condition is that the model sees actual system data, not a description of it. An assistant answering from general knowledge is useless in practice — and you can tell the difference by whether it can answer a question about a specific order.

Where AI does not pay off

  • Automatic repricing without bounds — the mistake shows up in revenue before anyone notices it.
  • Replies to complaints — they need judgement, and the customer can tell.
  • Generating content nobody reads — text produced only because it is possible is cost without return.
  • Deciding on cancellations or refunds — irreversible decisions belong to a person.

How to work out the real cost

Language models are billed per unit of text processed, not per user. The cost of translating one product description is measured in cents; so is the cost of one chatbot answer. For a thousand products in four languages that is tens of euros one-off, not thousands.

The problem starts with flat fees. A tool charging a fixed monthly amount for "AI" has no reason to show actual consumption — and you cannot tell whether you are paying for ten calls or ten thousand. Ask for a breakdown by module and by model.

How MitoOps handles it

The AI centre contains three separately licensable parts: an internal assistant over live system data, a website chatbot for shop visitors, and a copywriter for product text and translation. The customer chooses the model — with a fiftyfold spread in model prices, one shared model makes no sense.

Billing is on actual consumption, converted into credits: the rate belongs to the model, so a cheap call never subsidises an expensive one. The app shows usage broken down by module and by model, with a forecast of when a budget will run out. The system works normally without this module — AI is an option, not a requirement.

Common questions

Which model is best for an online shop?
It depends on the task. For chatbot answers, which are many and short, a cheaper and faster model fits. For a product description written once and living on the page for years, the best model is worth it. One shared model for everything always compromises one side or the other.
Does the model see my customers’ data?
Ask about this specifically. There is a difference between "the model receives a question and an answer from the database" and "the model receives the whole database". The second should not exist, and if it does, you want to know where the data goes and whether it is used for training.
Will AI replace customer support?
No. It takes over the repeated questions and frees time for the ones where a person cannot be replaced. A shop that uses AI to abolish support saves a salary and loses customers.
How should I start with AI in my shop?
With one task that has a measurable outcome — translating descriptions for one new market, for instance. If it pays off, you know why. Starting five things at once makes it impossible to tell which of them delivered anything.

Dealing with the same thing?

MitoOps brings stock, orders and shipping into one system. Let us walk through your numbers.