Digital Banking Conveyor: How BPM Connects People, Robots, and LLM Agents

Denis Ermilov Deputy Director for Strategic Clients at Cinimex Banking digitalization has entered a phase where the number of technologies is growing faster than the level of manageability. RPA, AI, and low-code accelerate individual operations but do not necessarily improve the overall outcome. Processes become fragmented, accountability becomes blurred, and control is often performed after the fact. In such an environment, a bank can paradoxically become more automated while simultaneously becoming more expensive to operate — simply because there is no unified management layer.

Imagine an airport: you can buy more aircraft and hire more pilots, but without air traffic control, procedures, queues, priorities, and safety rules, you will get chaos rather than increased capacity. The same logic applies to banking: robots and AI agents are the “aircraft,” teams are the “pilots,” and BPM is the “control tower” that connects everything into a single workflow and manages deadlines, exceptions, and accountability. Without the control tower, speeding up one part of the process simply creates a bottleneck elsewhere.

A common mistake is to build digitalization in a fragmented way: RPA separately, AI separately, and integrations separately. This leads to different task queues, different rules and reports, while the key question remains unanswered: who is responsible for the end-to-end outcome — a person, a robot, a system, or a business unit? At the pilot stage, everything may look promising, but at the scale of a bank, operational risks, maintenance costs, and the number of cases requiring manual intervention increase, while the benefits of automation stop scaling.

At one of Russia’s largest banks, this problem became particularly apparent: transaction volumes were growing faster than the capacity of operations teams, a significant share of back-office processes was still performed manually, and the IT landscape was fragmented. Task allocation and deadline monitoring relied heavily on manual management “dispatching,” meaning that business growth led to more delays, reprocessing, and costs.

The solution was not framed as a collection of individual automation projects, but as a transition to an industrial execution model — a digital conveyor where every operation becomes part of an end-to-end process, tasks are routed automatically, statuses and bottlenecks are visible, and accountability and control are established at the process level rather than across individual systems or teams.

In practice, this delivers the key benefit: more operations are completed without manual intervention, fewer cases require reprocessing or emergency intervention, deadlines are met more consistently, and the cost of execution no longer grows proportionally with transaction volumes. AI and robots are not deployed as isolated “features,” but as a way to systematically increase the share of automated operations within a controlled workflow.

One of the key lessons from such cases is that automation does not scale without orchestration. The bank implemented a multi-level model:

• STP automation for high-volume flows — where operations should be processed end-to-end according to predefined rules: standard payment and settlement operations, routine checks of account details and limits, scheduled reconciliations and end-of-day closing, and the generation of standard documents and notifications.
• RPA to eliminate tactical routine and integration gaps — for example, when manual actions in user interfaces need to be removed quickly or systems need to be connected without modifying their underlying applications.
• LLM agents for text and document processing — to accelerate the preparation of materials and improve information processing: initial classification of requests and documents; extraction of key facts from document packages; preparation of draft customer responses; case summaries for handovers between departments; draft explanations for compliance requests; and templates for internal memos and decision records.

Crucially, AI agents do not operate autonomously. They are embedded into BPM processes as controlled participants: they are invoked at a specific point in the workflow, receive structured context, and their output is recorded and reviewed. Responsibility for the final decision remains with a human, which is critical in a regulated banking environment.

From this perspective, BPM is not simply about process diagrams. It is the management layer of digital operations. As a corporate platform, KomandaPRO, based on Camunda, can serve as the core for process orchestration, execution control, and traceability.

In banking, the winner is not the organization that “implements the most AI,” but the one that makes AI and robots manageable participants in an end-to-end process.

The next step is to manage “digital employees” as a system. In the third article, we will examine a practical framework: what rules are needed for invoking robots and AI agents, how to establish quality control and auditing, where mandatory human checkpoints should be placed, how to define accountability and service-level agreements (SLAs) across end-to-end processes, and why this is particularly important in banking.

As an example, we will look at a “non-standard” operation involving front-office, back-office, compliance, and legal teams, and show how the digital conveyor can distribute work between people and digital employees without compromising control.