Frequently Asked Questions

An AI solution for business is a system that learns from a company's data and takes on a decision or an action that previously required a person: forecasting demand, scoring the risk of customer churn, classifying requests, choosing a recommendation, finding an answer inside corporate documents.

Artificial intelligence for business differs from ordinary automation in that it works from a pattern derived from historical data rather than from a rule written out in advance.

AI automation pays off fastest on repetitive operations that do not reduce to a simple rule: processing and checking documents, classifying and routing requests, preparing and normalising data, drafting routine texts.

Where a process is described by clear conditions, classic business process automation is cheaper and more reliable. We separate the two layers: AI takes on exactly the step that requires "understanding" a text, a document or a behaviour.

AI data analytics applies where the volume of information exceeds what can be reviewed by hand: finding patterns in transactions, behavioural clustering of the customer base, anomaly detection, testing business hypotheses against historical data.

On the Planeta Kino project, ML algorithms broke the base into behavioural clusters using more than a year of transactions, and the results were validated together with the marketing team through in-depth interviews.

Predictive analytics forecasts future events from historical data: demand, sales, capacity, stock, the probability of customer churn. Unlike ordinary reporting, which describes the past, forecasting analytics buys you time to act.

A practical example: for New Products Group an ensemble ML model forecasts production volumes with over 80% accuracy. The company moved from reacting to shortages or surpluses after the fact to planning volumes in advance.

Yes, and this is what makes a solution work in practice. AI integration runs through REST API, message queues, webhooks or a direct connection to the data warehouse — whichever suits your infrastructure.

An example from our work: at Planeta Kino the churn risk score is passed to Salesforce Marketing Cloud, and the customer receives a message matched to their risk level. Forecasts can reach an ERP for production planning or BI reporting in the same way.

Yes — developing AI solutions around a company's specific processes is our main format of work. Custom AI solutions are needed when an off-the-shelf service does not account for the specifics of the business.

A telling case is goodwine: the client planned to use a ready-made recommendation service, but testing showed the model did not take warehouse stock into account. Our own model, combining behavioural and logistics factors, delivered +5% average order value and +10% stock rotation.

You need a history of the events you want to forecast or classify: transactions, requests, orders, production figures, user behaviour. What matters is not only volume but also the depth of the period and how completely events were recorded.

The state of the data almost always turns out to be a task of its own. In the New Products Group case some retail outlets did not record zero stock, so the system confused an absence of sales with an absence of demand — approximation methods corrected that distortion and materially improved the forecast. This is why we start an AI rollout with a data audit rather than with choosing a model.

The timeline depends on the state of the data and on how many systems the integration touches. For reference, the real durations of our projects: the churn prediction model for Edenred took five months, the analytics and churn model for Planeta Kino eight months, and production forecasting for New Products Group and the recommendation system for goodwine twelve months each including rollout.

Testing a hypothesis on a PoC is usually much faster. We give a precise estimate after the data audit and once the scope of the first phase is agreed.

Yes, and for most tasks that is the most sensible start. A PoC checks one thing: whether the data holds a signal strong enough for the accuracy you need. That turns the decision about further investment into a matter of fact rather than expectation.

We apply this approach systematically. The first version of the model for New Products Group was built on a limited set of features while the tender was still running and, despite its simplicity, delivered the highest accuracy of all participants. The full ensemble model was developed only after the hypothesis was confirmed.

Yes. After launch a model needs support: forecast quality is monitored and the model itself is retrained on new data, because customer behaviour and the market change.

Scaling usually goes one of two ways: extending to adjacent tasks and categories — at New Products Group the first model became the foundation for other product lines — or connecting new channels and systems, as at Planeta Kino, where recommendations run simultaneously in push, email and the mobile app.