AI Agents in Business Processes
How a Low-Code Platform Ensures Control and Security
IT Alliance Solutions
How to integrate artificial intelligence into corporate processes without compromising governance, transparency, or security requirements.
Processes
BPMNAccess
RBAC · ABACAuthentication
KeycloakThe Challenge: AI Without Control Is a Risk
Banks and financial institutions are actively exploring the potential of large language models (LLMs). However, the transition from pilot projects to production deployment presents a systemic challenge: an autonomous AI agent can be unpredictable.
In a critical business process, an AI agent may generate a hallucination, access data directly while bypassing established access-control policies, or make a decision that cannot be adequately explained to a regulator.
01Risk
generate a hallucination
02Risk
access data directly while bypassing established access-control policies
03Risk
make a decision that cannot be adequately explained to a regulator
In the financial sector, this is unacceptable. The answer is not to abandon AI, but to embed it within a governed orchestration framework, where every step taken by the agent is controlled, validated, and auditable. This is the approach implemented in the ITA-FORMS low-code platform (Russian Register of Domestic Software No. 6468).
ITA-FORMS is not about “AI for AI’s sake.” Instead, it offers three practical scenarios in which an AI agent becomes an integral part of a proven BPMN-based orchestration framework with enterprise-grade controls.
- controlled
- validated
- auditable
Approach: Three Scenarios for Integrating AI Agents
2. Platform as a Tool Registry
In this scenario, the AI agent acts as a coordinator, while the platform exposes its existing system integrations through a standardized protocol, such as OpenAPI or the Model Context Protocol (MCP).
Connectors to ERP and CRM systems, payment gateways, and databases that are already configured in the Low-Code Designer are automatically exposed as a set of functions available to the agent through Tool Calling. There is no need to build separate integrations for AI—the platform makes its existing connectors available to the agent.
The key principle is no direct access. The agent never accesses a database or external API directly. All calls are routed through the platform’s integration bus and subject to:
RBAC/ABAC
Access control (RBAC/ABAC) — the same policies used by other system components;
Logging
Logging of every request for audit purposes;
Rate limiting
Rate limiting to protect backend systems from overload.
Scheme
Platform as a Tool Registry
Principle
Integrating AI does not introduce new access paths:
the agent operates under the same access-control policies as any other platform component, while existing integrations can be reused without duplication.
Security and Deployment
All three scenarios are built on the core architectural principles of ITA-FORMS:
On-premises deployment.
AI agents operate within the customer’s environment. Data remains within the organization’s security perimeter, which is critical for financial institutions and regulatory compliance requirements.
Unified IAM infrastructure.
Existing authentication and authorization mechanisms, including Keycloak, ABAC, and RBAC, are used. There is no need to establish a separate identity and access management system for AI.
End-to-end traceability.
Every agent action is recorded in the audit log alongside other process steps, providing a complete and auditable trail of AI activity.
Microservice architecture.
AI components are isolated from the platform core. Updating or replacing an AI model does not affect running business processes.