Mass implementation of AI in the processes of the logistics company. How forecasting, LLM and RAG became part of everyday work
17 March 2026
Introduction: why PEC needed AI
A major logistics company with a distributed network across multiple locations and a complex IT landscape. The company has been operating since 2001, and analytical and forecasting tools here began to be developed long before the AI boom. According to Anton Sluchenkov, Director of Development and Data Visualization Department at a major logistics company, it was the growth in process complexity and the number of metrics that became the key trigger for moving to ML and LLM.
Initially, analytics were built for top management, but over time they filtered down to branches, warehouses and even individual employees. At the same time, the number of forecasted indicators increased – from total transportation volumes to the consumption of specific materials. At some point it became obvious: traditional tools no longer scale.
From Excel to ML: how industrial forecasting was built
Historically, forecasts at the company were made in Excel and then moved to SQL. Formally it looked like automation, but in fact it remained the same Excel, only “rewritten in code”. Each indicator required its own approach, formulas became more complex, and maintenance became increasingly painful. The situation worsened when one of the key specialists, who effectively held the forecasting logic, left the company.
The solution was a switch to machine learning. The company built a unified ML forecasting pipeline: data are cleaned and normalized, features are automatically generated – seasonality, lags, moving averages, calendar factors. The main production model used is ARIMA, and for some metrics – gradient boosting. A fundamental point: there is no "magic model for everything" here. For each metric the system itself chooses the most suitable approach based on quality metrics.
Forecast quality is regularly monitored: forecasts are compared with actuals three times a month, backtesting and degradation monitoring are performed. As a result, the average forecasting accuracy today is about 95%, and the number of supported metrics has grown from dozens to hundreds. Adding a new indicator no longer requires rewriting code – configuration is sufficient.
LLM in telephony: how calls started filling CRM automatically
The next large case was the implementation of LLM in telephony. The company has accumulated thousands of customer call recordings, but managers were reluctant to fill in CRM: there was simply not enough time. The task was simple – automatically extract the meaning of the call and turn it into structured data.
Technically, the process looks like this: calls are saved in WAV format and then transcribed using a specialized speech-to-text solution. Recognition quality is production-level, with WER under 5%, which allows confident work with text. Then GPT OSS 20B comes into play, which determines the call topic, key meanings and client intents.
A separate challenge was topic classification. PEC uses about 180 fixed topics, and direct determination of a single topic works poorly. Therefore a two-step approach was applied: first the model selects three most suitable topics, and then it refines the final option. A vector store is used for this – without it accuracy drops sharply.
The result is automatic filling of the client card in CRM: purpose of the call, agreements, next step, product of interest. The manager only needs to verify the data, not fill everything from scratch.
