Gazprombank.Tech shows how AI-personalized debt collection raised recoveries by 25%

Gazprombank.Tech described how it rolled out Next Best Action for handling overdue debt. The team combined a new data architecture with uplift models to pick the best contact channel for each client and recover money faster, without adding load on the call center.

Where the days were lost

Before the new setup, the bank followed the logic of a classic data platform: sources were replicated into an operational store, passed through the Stage and Core layers of the corporate warehouse, and then landed in Hadoop marts for model training and scoring. The chain was reliable but too slow for collection. The most sensitive information, the fact that a payment was missed, reached the models with a lag of up to three days. The client had already failed to pay, yet the bank's personalized response did not start until the day after tomorrow.

On top of that sat the old communication logic. Contracts were sorted into a Balance at Risk matrix: the higher the risk and balance, the more contacts a client received. Then came a fixed cascade of actions: push, SMS, robot call, operator call. This approach worked at the segment level, but it did not answer the more important question: which channel will work for this particular client right now. As a result, some people were bothered needlessly, and the call center spent resources on calls that could have been skipped.

How NBA was built

The team chose not to break the heavily loaded Collection business system or add another heavy process to it. Instead, the developers found an existing automated job that fires the moment a delinquency appears and hooked into it through the ETL Framework. The relevant events were tagged and immediately exported to Hadoop and Data Factory, bypassing unnecessary delays. As a result, the bank began learning about a delinquency almost the moment it occurred, without creating extra load on the core system.

The team then changed the decision model itself. Instead of segmentation, it moved to First Best Action and Next Best Action: the system does not just assess risk, it recommends the best next contact channel. For the pilot, the team tested several uplift approaches: solo models, independent and dependent pairs of models, and multiclass and multi-treatment variants. They were judged not on a single attractive metric but against a set of constraints at once: Gini, Balance at Risk, Qini, communication costs, budget and call-center operator availability.

"We decided not to choose anything at all and just write an optimizer."

The result is an optimizer that, at each stage of delinquency, trains and tunes a set of models and then picks the best option under specific business constraints. At the early First Best Action stage, the system directly recommends whether to write to the client, send a push or SMS, or hand the case to a robot or an operator. The bank also maintains quality control models, so it can see not only the final conversion but also how the uplift models behave within each group.

What the pilot showed

The pilot was designed as a fair experiment. About 30% of the flow went to the test loop, and the rest was split between the old strategy and the new recommendation system. Data collection began in late 2024, followed by another two months of development and rollout to production. According to the team, the project fit within a normal model lifecycle but produced a noticeable effect already at the first stage of collection.

• The settlement rate rose 25% relative to the current strategy.
• For individual communication channels, the effect delivered 15% to 32% in additional profit.
• Operator calls fell 73%, and the cost of communication per client fell 10%.
• Client response rose 5% for operator calls and 11% for robot calls.
• The financial effect exceeded 3 billion rubles in preserved balances per 20 thousand contracts in just one month at the First Best Action stage.

What this means

The Gazprombank.Tech case shows that in sensitive processes like collection, the winner is not the most aggressive communication but the most precise. If you remove the data lag and choose the contact channel for each specific client, AI can raise recoveries, cut costs and avoid burning customer loyalty with needless calls.