Digital Employees in Banking: Managing Them Through BPM — Rules, Control, and Audit

Denis Ermilov Deputy Director for Strategic Clients at Cinimex In the previous article, we showed how BPM connects people, robots, and AI agents into a single process. Now we turn to the question of how to manage this environment. “Digital employees” — RPA robots and large language model (LLM)-based agents — have already become a real resource for banks: they take over routine tasks, speed up processing, and help handle peaks in workload. But the more such digital workers a bank has, the more important management becomes: who is responsible for a robot’s actions, what rules does it follow, where is the result recorded, and how can the correctness of a decision be demonstrated during an audit or incident investigation? Therefore, the next step after implementation is not “more automation,” but manageability: clear rules, measurable performance, control, and audit.

Imagine an industrial factory: you can install robots and speed up the production line, but without process specifications, dispatching, quality control, and production logs, you will get not higher output but more defects and downtime. Banking operations work the same way: robots and AI agents are the “machines,” people are the “operators,” and the management framework — process orchestration and control — is the production line that defines the sequence of operations, checkpoints, and accountability. It is this framework that turns disparate automation initiatives into predictable results.

The most dangerous mistake is to treat digital employees as “utilities”: each team launches its own robots, writes its own scripts, and connects data “as needed.” This may look effective during pilots, but at a bank-wide scale it creates an uncontrolled execution layer: different versions of logic, inconsistent rules, no unified action log, and increasingly complex compliance and audit requirements. As a result, the bank may have many automation solutions but little manageable productivity.

One of Russia’s largest banks has already gone through this transformation. Before the change, transaction volumes were growing faster than the capacity of the operations function, a significant share of operations was performed manually, and queue and deadline management relied heavily on “manual dispatching.” The solution was not a collection of individual robots, but the creation of an industrial operating model — a digital “production line” for back-office operations: a unified task flow, algorithmic routing based on priorities, deadlines and competencies, real-time quality and service-level monitoring, and people and digital employees operating as a single resource pool. This approach delivered measurable results: productivity increased 2.5 times, the cost of the operations function fell by more than 20%, and business growth no longer required a proportional increase in operating costs.

Managing digital employees in a bank should be as rigorous as managing people and critical systems — and should be based on measurable performance. In practice, this means three levels of discipline.

The first is governance and accountability: a “passport” for every robot or AI agent specifying its purpose, owner, scope of authority, data sources, and failure scenarios.

The second is control and traceability: embedding the digital employee into a process as a managed participant — invoked at a strictly defined point, provided with structured context, producing a recorded result, and subject to mandatory human review wherever the decision affects risk or the customer. This also requires full traceability and auditability: who did what and when, which version of the logic or configuration was used, what input data was provided, and what output was generated. Change management is equally important, including version control, regression testing, and approvals.

The third is quality and security: access rights based on the principle of least privilege, risk controls built into the process — for example, four-eyes approval, selective review, correction and reprocessing procedures — as well as a regulated mechanism for stopping automation and returning to manual processing in the event of an incident.

Performance indicators should reflect business impact rather than the “number of robots”: the share of straight-through processing (STP), the share of manual operations, turnaround time (TAT), compliance with service-level agreements (SLA), reprocessing and error rates, resource utilization, queue length, and unit cost per operation.

For example, in a “non-standard” operation, an AI agent can prepare a case summary and draft explanations, while a robot can collect data from multiple systems. The process itself records the input data, configuration version, and result, and then sends the case to the responsible employee for mandatory approval. In the case described above, this approach made it possible to achieve 96% end-to-end automation across key workflows involving 5.4 billion transactions per year, reduce manual work by 90%, and deploy a “digital workforce” of more than 200 robots as a managed operational resource — supported by a real-time performance dashboard and KPI tree.

This type of management model is typically built around BPM orchestration. One example is KomandaPRO, a platform based on Camunda that makes it possible to establish execution rules, control mechanisms, and auditability within a single framework.

In the next article, we will break this down into a practical “management toolkit”: what artifacts are required — a digital employee registry, employee passport, competency matrix, control rules, and audit log — how to manage operations on a daily basis using data — a performance dashboard, KPI tree, and deviation analysis — and how to properly distribute work between people, robots, and AI agents in a complex banking scenario involving the front office, back office, compliance, and legal teams. The goal is to create not a collection of pilots, but a managed workflow with clear accountability and demonstrable results.