Gazprombank.Tech showed how AI personalization in debt collection increased debt recovery by 25%

Gazprombank.Tech shared its implementation of Next Best Action for managing overdue debt. The team combined a new data architecture with uplift models to select the optimal communication channel for each client and accelerate money recovery without overloading the call center.

Where days were lost

Before launching the new scheme, the bank operated under classic data platform logic: sources were replicated into an operational data store, then passed through Stage and Core corporate warehouse layers, and finally arrived at Hadoop data marts for model training and scoring. This pipeline was reliable but too slow for debt collection. The most critical information — the very fact of an overdue payment — reached the models with a lag of up to three days.

By then the customer had already missed their payment, yet the bank's personalized response would only trigger the day after tomorrow. On top of this came outdated communication logic. Contracts were placed into a Balance at Risk matrix: the higher the risk and balance, the more contact attempts the customer received.

What followed was a fixed cascade of actions — push notification, SMS, robocall, operator call. This approach worked for segments but failed to answer the key question: which channel will actually work for this specific client right now? As a result, some people were contacted unnecessarily, and the call center spent resources on calls that could have been avoided.

How NBA was built

The team chose not to overhaul the heavily loaded Collection business system by adding another heavy process. Instead, developers found an existing automated workflow that triggers when an overdue payment occurs and connected to it via ETL Framework. Relevant events were tagged and immediately exported to Hadoop and Data Factory, bypassing unnecessary delays.

This allowed the bank to learn about an overdue payment virtually at the moment it occurs without creating additional burden on the main system. After this, they transformed the decision-making model itself. Instead of segmentation, the team moved to First Best Action and Next Best Action: the system not only determines risk but recommends the optimal next communication channel.

For the pilot, they tested several uplift approaches — single models, independent and dependent model pairs, multi-class and multi-treatment variants. They evaluated them not by a single attractive metric but immediately against a set of constraints: Gini, Balance at Risk, Qini, communication costs, budget, and call center operator availability.

"We decided we wouldn't choose anything at all — we'd just write an optimizer."

In the end, an optimizer was created that at each stage of delinquency trains and tunes a set of models, then selects the best option under specific business constraints. At the early stage, the First Best Action system directly recommends whether to write to the client, send a push notification, SMS, transfer the case to a robot or to an operator. Additionally, the bank maintains control quality models to see not only overall conversion but also how uplift models behave within each group.

What the pilot showed

The pilot was structured as a fair experiment. Approximately 30% of the flow was directed to the test circuit, while the remaining volume was split between the old strategy and the new recommendation system. Data collection began in late 2024, followed by two months of development and production deployment. According to the team, the project fit within a normal model lifecycle but delivered noticeable results already at the first debt collection stage.

• Settlement rate increased by 25% relative to the current strategy.
• Individual communication channels showed effects ranging from 15% to 32% additional profit.
• Operator calls were reduced by 73%, and communication cost per customer dropped by 10%.
• Client response rate increased by 5% for operator calls and 11% for robocalls.
• Financial effect exceeded 3 billion rubles in saved balances per 20,000 contracts in just one month at the First Best Action stage.

What this means

The Gazprombank.Tech case demonstrates that in sensitive processes like debt collection, the most precise communication wins, not the most aggressive. If you eliminate data lag and select the contact channel for each specific customer, AI can simultaneously increase money recovery, reduce costs, and avoid burning customer loyalty with unnecessary calls.