The sudden rise of artificial intelligence has broken the traditional boundaries of data management. For years, data protection meant backing up files, setting permissions, and ensuring you could restore a virtual machine or a database if something went wrong. When you introduce large language models, retrieval-augmented generation pipelines, and autonomous agents into your environment, the risk profile shifts completely. Systems can remain fully operational from an infrastructure perspective while becoming completely untrustworthy because the data feeding them has been compromised, exposed, or corrupted.
During a presentation at AI Infrastructure Field Day 5, the team from Commvault laid out a fresh approach to this dilemma. The core of the problem is that artificial intelligence model deployment moves much faster than traditional infrastructure refresh cycles. Organizations are rushing to feed enterprise data into large language models to build chatbots, internal search tools, and automated workflows. If you do not govern that data before it enters the AI ecosystem, you are creating massive security and compliance liabilities. You cannot simply watch what the AI does. You have to control what the AI sees.
Jose Gomez, who joined Commvault with the acquisition of Satori, detailed how the company is tackling this challenge through a newly designed governance control plane. The goal is simple, but execution requires deep integration. You must manage access to structured and unstructured data across diverse environments like SQL Server, Snowflake, and Databricks before that information ever reaches an AI model or a human prompt. It represents an architectural shift from historical data security strategies. Instead of relying on static access control lists tied to specific storage volumes, organizations need real-time, dynamic intervention.
This new governance approach relies on a proxy-based application that acts as a Data Access Controller and an LLM gateway. Think of it as an intelligent checkpoint positioned right in the data path. When a user or an autonomous AI agent requests information, the proxy intercepts the query in real time. It evaluates the request against corporate policies, verifies user identity via standard identity providers such as Active Directory, and applies granular security rules. This happens instantly, ensuring that sensitive data is either redacted or blocked before it leaves its source repository.
The introduction of autonomous agents makes this real-time proxy architecture critical. Traditional security models assume a human is always sitting at the keyboard, making explicit decisions. AI agents do not work that way. They operate independently, moving across systems, querying databases, and synthesizing information at a pace humans cannot match. If an agent has over-privileged access, it can inadvertently pull sensitive financial records or personally identifiable information into a vector database used for retrieval-augmented generation. Once that sensitive data is embedded into a public or shared organizational model, clawing it back is incredibly difficult. The proxy model prevents this entirely by enforcing boundaries on the agent itself.
Data classification and automated redaction form the operational backbone of this control plane. When the proxy intercepts data destined for an AI pipeline, it scans for sensitive information patterns, such as social security numbers, credit card details, or proprietary code blocks. The system can automatically mask or strip these elements out on the fly. This means you can still use your operational data to train models or inform downstream applications without the fear of leaking regulated information into training logs or public contexts. It provides a way for enterprises to innovate without compromising security.
Sovereignty and regional compliance add another layer of complexity to modern AI infrastructure. Regulations like GDPR impose strict rules on where personal data can be processed and stored. If your AI workloads are running in a public cloud region across the world, you might be violating compliance laws just by running an inference query. Commvault addresses this by offering flexible deployment options for the Data Access Controller. Organizations can choose between a fully managed cloud service or a customer-hosted Kubernetes cluster. By hosting the control plane locally on Kubernetes, enterprises keep data processing within specified geographic boundaries, satisfying sovereignty mandates while utilizing global AI services.
Security architecture is never a single product fix; it is a continuous alignment of capabilities. True resilience means acknowledging that governance and recovery are two sides of the same coin. If a data breach or a rogue model configuration occurs, controlling access helps contain the blast radius, but you still need a trusted path back to a known-good state. Traditional recovery tools focus on isolated virtual machines, but AI systems are webbed with interdependencies among models, pipelines, vector indexes, and identity registries. The entire stack must be protected holistically so that the business can restore a coherent state when trust is lost.
The strategy presented indicates that the conversation around artificial intelligence and data is shifting away from simple metrics such as speed and capacity. Feeds and speeds matter, sure, but they are secondary to control, visibility, and safety. Organizations do not just need faster storage pipelines; they need unified data platforms that build guardrails around training and inference workflows. This control plane serves as that critical safety net. It allows IT leaders and security teams to collaborate, closing the historical ownership gap that often leaves AI deployments unprotected.
Enterprise technology is messy, and implementing guardrails around fast-moving projects is always an uphill battle. The key is to embed governance directly into the infrastructure layer rather than treating it as an afterthought. By using a proxy-based control plane, companies can enable their development teams to experiment with external harnesses such as Claude or Copilot while maintaining absolute control over the data foundation. It balances business efficiency with robust compliance.
To see the full technical breakdown and demonstrations of this architecture, you can review the session recordings on the appearance page. For broader coverage of the technologies shaping enterprise technology deployments, visit TechFieldDay.com.