Today, the fascination with ad-hoc experiments using large language models (LLM) is giving way to pragmatic calculation. Organizations are shifting from chaotic AI usage to building managed AI processes. The main challenge for CIOs and Chief Risk Officers (CROs) is to make security and compliance an organic part of the enterprise IT architecture, rather than an additional bureaucratic burden.
According to research, about 13% of organizations—so-called "Pacesetters"—consistently outperform others in generating real value from AI implementation. The primary reason for their success is that they integrated risk management into the foundation of their systems from the start.
Most enterprise system architects face a serious threat: "Shadow AI" and uncontrolled data leaks. When models are integrated into business processes chaotically, without proper risk management and auditing, the company is at risk. The problem is that AI models are often implemented as isolated tools, whereas they should be part of an orchestrated business process where every step is logged and data access is strictly controlled at the infrastructure level.
Why chaotic AI becomes a threat: from "shadow" prompts to OWASP vulnerabilities
The OWASP organization, in its current Top 10 Risk & Mitigations for LLMs and Gen AI Apps, identifies indirect prompt injection (vulnerability LLM01:2025) as the primary and most critical risk for LLM applications. An attacker can manipulate the model's behavior through business process input data (e.g., text in an incoming email or other document), forcing the AI to perform unauthorized actions.
Additionally, there is a constant risk of sensitive information leakage. If a model is integrated without proper access control, it may generate responses based on data that the current user is not authorized to see. OWASP highlights the persistent nature of these risks, which requires continuous monitoring rather than a one-time technical fix.
Security framework: adapting NIST AI RMF 1.0 for corporate BPM
To structure and address these issues, the NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0) was developed. It organizes AI risk management around four key functions. Applying these requirements to the executable process architecture looks as follows:
- Govern: Requires clear separation of roles and access to data used by AI. Security policies must be embedded in the platform infrastructure.
- Map: Mandates clear documentation of where the model is called within a process, what data it receives as input, and where it sends the result.
- Measure: Involves risk assessment, monitoring model decisions, and logging events in real-time to detect anomalies.
- Manage: Requires rapid response mechanisms, including the ability to instantly change business rules without stopping or refactoring the entire system code.
Process orchestration using BPMN 2.0 and DMN
A reliable way to integrate AI into corporate processes is by using the executable BPMN 2.0 standard and the DMN business rule description standard. BPMN 2.0 allows for building a model that serves as both documentation and a process management tool, providing visibility into the state of each instance.
- Separation of business logic (DMN) and process flow (BPMN): Instead of hardcoding decision-making rules into the AI model, strict constraints are moved to DMN tables. This allows for rapid updates to AI operational rules.
- Configuring audit trails in the BPM system: To comply with NIST requirements, every step performed by AI must be recorded in the system audit log. This is critical for reconstructing the decision chain during deviation analysis.
- Implementing human-in-the-loop verification: Process routing allows for automatically directing questionable AI decisions to a competent employee for verification.
Access control and audit trail: implementing requirements on the UnityBase platform
It is important to remember that implementing automation tools alone does not guarantee automatic compliance with NIST AI RMF—it is merely the technical foundation for meeting the framework's requirements. However, having the right platform significantly simplifies the technical architecture of compliance.
For building enterprise solutions, Intecracy Group—an alliance of independent companies linked by partner agreements and share exchanges—uses the low-code platform UnityBase (a joint development of the consortium companies, where InBase is a key developer). Products built on this platform, such as Scriptum (BPM/DMS), use built-in security mechanisms for AI risk management.
Row-Level Security (RLS) and Access Control Lists (ACL) mechanisms, implemented at the platform level, ensure that the AI model can only interact with data for which the process initiator has permissions during a request. This minimizes the risk of data leakage via prompt injection. A detailed system audit trail records the model's actions, ensuring accountability. The platform also supports independent file stores (BLOB/file stores), allowing documents to be isolated from the general data pool.
Process mining as a tool for detecting anomalies in AI agent performance
Even with well-designed BPMN models, deviations can occur in real-world scenarios. Process mining technology allows for identifying how a process is actually executed based on system logs, helping to find bottlenecks and "shadow" routes.
In the context of AI, process mining is used to detect whether AI agents are creating hidden processes that bypass corporate security policies. For example, detecting overly frequent or non-standard requests from an automated agent to external systems allows for timely adjustment of security parameters.
| NIST AI RMF Function | Process Requirement | Technical implementation in BPM (based on UnityBase / Scriptum) |
|---|---|---|
| Govern | Creation of access policies and roles | Use of RLS (Row-Level Security) and ACL mechanisms that limit the volume of data for AI requests. |
| Map | Visualization and documentation of data flows | Use of the BPMN 2.0 standard for clear definition of AI model entry and exit points in the process. |
| Measure | Risk assessment and decision monitoring | Logging every AI step in the system Audit Trail for deviation analysis and prompt injection detection. |
| Manage | Rapid response and logic modification | Separation of business rules (DMN) and flow (BPMN) for instant adjustment of AI behavior without changing platform code. |
FAQ
How to integrate NIST AI RMF into existing company business processes?
Integration begins with mapping (Map): it is necessary to identify and document all AI interaction points in BPMN models. Then, at the architecture level, data access rules are implemented, and a mandatory system audit trail is configured to record every model decision.
Which OWASP security risks are most critical when working with LLMs in BPM?
The most critical are indirect prompt injection and the risk of sensitive data leakage. These are mitigated by restricting model access rights at the database level and implementing continuous monitoring.
How to ensure the reliability of logs for auditing decisions made by artificial intelligence?
This is achieved using the system audit trail tools of the platform where the solution is deployed. This allows for recording transactions and metadata changes, ensuring the ability to reconstruct the AI decision-making chain at any time.