Organizations are shifting en masse from experimenting with generative AI to implementing autonomous AI agents (agentic workflows). However, IT departments face significant challenges: integrating these intelligent entities into existing corporate environments often carries risks to security, budget predictability, and data management. According to the Cisco AI Readiness Index 2025, only about 13% of organizations are classified as fully prepared ("Pacesetters") to scale such technologies and derive stable value from them.
The remaining companies face infrastructure limitations. Success in deploying autonomous AI agents depends not on the choice of a specific large language model (LLM), but on the fundamental readiness of the IT infrastructure: applying FinOps cost controls, granular data access segregation, and protection against new attack vectors.
The AI readiness paradox: why model selection is secondary to infrastructure
When IT directors and architects plan AI integration, the focus often shifts to comparing model accuracy. However, in real enterprise systems, the model is merely a computational engine. AI agents act autonomously, and this autonomy creates an infrastructure challenge. An agent independently determines the sequence of actions to complete a task and can generate dozens of unpredictable requests to databases or external APIs.
Without strict infrastructure guardrails, there is a risk of system failures and data leaks. Attempts to solve control issues solely at the level of system prompts are ineffective due to vulnerability to Prompt Injection techniques. Security must be ensured at the level of network gateways, data platforms, and centralized policies.
FinOps for AI agents: preventing budget surprises from autonomous cycles
The autonomy of AI agents poses a direct threat to IT budget stability. A real-world example is uncontrolled cyclical API calls from agents, leading to sudden and unexpected spikes in cloud computing costs. If an agent encounters a formatting error in a database response and attempts to fix it independently, it may initiate a continuous loop of requests to paid APIs.
To avoid such situations, implementing the FinOps methodology is necessary. According to the FinOps Foundation, cloud cost management should be a shared responsibility between engineering, finance, and business teams. Microsoft Azure Well-Architected principles (Cost Optimization module) also provide for the creation of robust cost forecasting models and resource management for AI. Technical control should include:
- Dynamic resource quoting and guardrails on the number of tokens or request costs for specific agents.
- Use of API gateways to monitor traffic volumes and automatically block anomalous cycles.
- Detailed cost attribution for transparent allocation of compute consumption across business processes.
Data-level access control: why traditional RBAC is no longer sufficient
Another risk is unauthorized access by AI agents to sensitive datasets due to incorrectly configured Role-Based Access Control (RBAC). When an agent acts on behalf of a user or system, standard role-based access may allow it to read data across an entire department, leading to the compromise of confidential information.
To avoid this, an architecture supporting Row-Level Security (RLS) and Access Control Lists (ACL) is required. In this context, the use of the UnityBase platform—a full-stack JavaScript low-code framework developed by the Intecracy Group alliance (where InBase is a key, but not the only, developer)—is illustrative.
UnityBase utilizes a metadata-driven architecture. In the commercial editions of the platform (Enterprise and Defence), RLS and ACL mechanisms are built-in. When an AI agent makes a request via the platform's automatically generated REST API, the system processes the request within the context of a specific user's session. The agent is physically unable to extract records from the database that the user does not have rights to, which eliminates the risk of data leakage at the infrastructure level.
AI Fortification: protecting infrastructure from autonomous agent compromise
Deploying autonomous agents also requires preparation for new cybersecurity threats. According to the Cisco Cybersecurity Readiness Index 2025, an organization's modern readiness to repel cyberattacks is assessed using 31 technical solutions across 5 key pillars. One of these pillars is AI Fortification—protecting the model itself, as well as its input and output streams, from compromise.
For the safe operation of agents, isolation of their execution environments (e.g., running custom scripts in secure containers) and anomaly monitoring are necessary. Infrastructure protection involves mandatory sanitization of input data to mitigate Prompt Injection risks, as well as robust auditing of agent actions with full logging of its requests and database changes.
Architectural foundation: building a secure AI environment without "gray zones"
One of the main challenges for the enterprise segment is the emergence of "Shadow AI"—the deployment of agents by business units bypassing security and data governance policies. Creating a centralized architectural foundation helps avoid this.
The Intecracy Group alliance offers a comprehensive approach to solving this task, combining the expertise of Softengi in developing custom AI-native systems (Softengi is certified to the international AI management standard ISO/IEC 42001:2023) with the platform mechanisms of UnityBase. Based on this platform, enterprise products with built-in AI scenarios are developed, such as the Megapolis.DocNet electronic document management system (which holds a G2-level information protection certificate and has its own AI center) and Scriptum.DMS.
These solutions leverage the infrastructure advantages of UnityBase: a unified domain metadata model, strict auditing, and RLS. This approach ensures that AI agents function exclusively within the permitted security perimeter, preventing uncontrolled access to corporate content and ensuring the business's architectural readiness for AI scaling.
| Infrastructure component | Risk without preparation | Technical solution |
|---|---|---|
| Cost control (FinOps) | Uncontrolled costs due to cyclical API requests from autonomous agents. | Implementation of dynamic guardrails at the API gateway level, resource quoting per the FinOps Framework. |
| Data access (Data Governance) | Unauthorized agent access to sensitive data (HR, finance). | Application of Row-Level Security (RLS) and ACL at the data platform level (e.g., via UnityBase mechanisms) to isolate context. |
| Security (AI Fortification) | Injection of malicious instructions (Prompt Injection), data leakage. | Assessment via 31 Cisco Cybersecurity Readiness solutions, isolation of agent execution environments, request logging. |
FAQ
How does FinOps help control costs for LLMs and AI agents in the cloud?
Implementing the FinOps methodology creates a shared responsibility model for cloud resource usage. Technically, this is realized through dynamic quoting, guardrails on the number of tokens or request costs for a specific agent, and blocking anomalous cyclical calls at the API gateway level that could cause cost spikes.
What data security requirements arise when connecting AI agents to corporate databases?
Classic Role-Based Access Control (RBAC) is no longer sufficient, as an AI agent can autonomously generate requests. It is necessary to ensure granular access segregation directly at the database level using Row-Level Security (RLS) and Access Control Lists (ACL) so that the agent only accesses information permitted to the specific user.
How can Row-Level Security (RLS) be implemented to restrict AI agent access to documents?
Implementing RLS is possible at the architectural foundation level. For example, the low-code platform UnityBase in its commercial editions (Enterprise/Defence) has built-in support for RLS and ACL. Thanks to a metadata-driven architecture, all AI agent requests via the generated REST API pass through these filters within the user's session context, making data leakage impossible.