Data Governance for Agentic AI: How CISO’s Can Ensure Security and Compliance

Agentic AI is rapidly transforming how businesses operate, and with this evolution comes a critical need to rethink data governance strategies that will empower users to build and deploy secure, governed AI agents with confidence. 

In this Q&A, Chandrika Shankarnarayan, VP of Product and a Data Cloud expert at Salesforce, shares her perspective on the governance challenges CISOs and IT leaders face and the opportunities presented by agentic AI‌. She also shares how solutions like Data Cloud are helping organizations deploy cohesive governance offerings.

Data Cloud is the intelligent activation layer for all enterprise data, designed to coexist with data silos, enabling organizations to unify and activate real-time enterprise data bi-directionally and securely, no matter where it lives.

Q. What do you foresee will be the biggest data governance challenges as agentic AI replaces more traditional AI systems?

The role of employees is evolving. We’re moving beyond traditional management of other humans toward a future where collaboration between humans and AI agents in solving customer problems will be paramount. This shift necessitates seamless access to a wide range of data. This data includes structured information held in databases, as well as unstructured content like videos, audio, and text gathered from various internal systems and even publicly available resources.

The paramount challenge in this agent-driven future is governing this complex data landscape with privacy and security in mind.

Each of these data types and sources has its own governance policies, often managed across different platforms. The paramount challenge in this agent-driven future is governing this complex data landscape with privacy and security in mind.

Q. How is agentic AI forcing companies to rethink data governance, and what new tools are critical for success?

The fundamental reality of data existing in different systems and silos with varying governance doesn’t change. However, to make the agentic era work effectively ‌— where AI and human agents collaborate to solve problems accurately and quickly ‌— ‌we need a governance framework built on the concept of data fluidity. 

This is precisely what we’re doing with Data Cloud and the Salesforce Platform. Data lives everywhere, and we don’t want to force it into a central location. The key is to provide a framework that allows us to look at data and metadata holistically, with common classifications and taxonomies. Having a unified framework like this enables us to create consistent policies around metadata, regardless of where the data lives. The goal is to provide agents and humans with easy, consistent, and quick access to the data they need to reason, query, and provide answers.

Data Cloud harmonizes data from all sources (ingested in Data Cloud or federated through Zero Copy) into a single, real-time customer profile, enabling AI-driven insights and actions. By breaking down silos and integrating seamlessly with Salesforce’s ecosystem, we ensure businesses can leverage their data to deliver personalized experiences, optimize operations, and drive growth.

Q. What challenges does agentic AI that operates more autonomously and interacts with data in new ways, present for maintaining data quality, integrity, and lineage?

This is a crucial point. For agentic AI to provide efficient and accurate responses, especially when dealing with both Salesforce and non-Salesforce data, understanding the additional metadata context is vital. 

We have a unique advantage here, understanding the context of sales, service, marketing, and various industries. By providing agents with this rich business and technical contextual data alongside data transformation and management context, we can ensure better outcomes. Furthermore, it’s a continuous lifecycle: We create the context, lineage, and metadata catalog, feed it to the agent, and the agent’s responses then continually enrich the metadata, contributing to accuracy and speed.

Q. How do you anticipate agentic AI’s increased autonomy impacting data privacy regulations like GDPR or CCPA? Are new mechanisms needed for ongoing compliance?

Data privacy and compliance have been long-standing concerns. In the agentic era, the key difference is agents’ ability to access and reason over data within defined guardrails and interact with humans who traditionally performed this work. The crucial aspect is enabling agents to access data and take actions in a governed fashion, adhering to policies set up for their specific roles. 

With Data Cloud, for example, data tagging and classification allow us to label data according to various needs like GDPR, CCPA, and quality. This classification then underpins a unified policy framework for access, masking, data purpose, and privacy. By applying these policies to both human and AI agents based on their roles and corresponding permissions, we can ensure uniformity and maintain trust in meeting compliance requirements.

Q. Are there any new features within Data Cloud that will change data governance significantly in the age of agentic AI?

Absolutely. At Salesforce, we’re focusing on several key areas. 

Firstly, the ability to easily govern both structured and, crucially, unstructured data. Our new capabilities allow for automated tagging and classification of all data, including structured and unstructured data, enabling policy-based governance across various use cases over all data – both data that is ingested into Data Cloud as well as federated (Zero-Copied) and living in other systems in the customer’s ecosystem. We have seamless metadata integration that makes true enterprise governance a reality for our customers in this agentic era. 

Secondly, we’re empowering users to author and manage these policies through a unified framework based on automated tagging and classification, enforcing them consistently regardless of where the data resides. 

Finally, in the realm of security and privacy, we’re enhancing capabilities to prevent vulnerabilities and protect sensitive information, offering flexibility over encryption key management and ensuring data behind firewalls stays protected with Private Connect, which now applies to many data sources that we already support.

Q. How does the ability to handle unstructured data help with compliance, particularly with regulations like GDPR?

Consider a financial institution. They handle structured data like credit card transactions, but also unstructured data like bank policies in PDFs or customer communications. When an agent interacts with this data, there needs to be rules and policies governing what they can access and do. 

For example, an AI agent shouldn’t automatically approve a credit limit increase, but a human agent reviewing the request needs access to all relevant data, including company policies, while still adhering to privacy rules. This is where governing unstructured data becomes critical. We can set up privacy / compliance policies to ensure that agents, human or AI, can access the necessary information to complete a workflow while still masking sensitive data like credit card numbers, social security numbers, etc. This ensures compliance even when dealing with diverse data types.

Q. Do you see privacy-related acts like GDPR evolving to account for agents, or will the principles remain the same?

In the near term, I don’t foresee these regulations drastically changing specifically for agents. These rules are deeply ingrained in numerous systems in the industry. Furthermore, AI agent frameworks are still evolving, with ongoing work to address hallucinations and inaccuracies. Therefore, a sudden loosening of compliance policies due to the rise of agentic AI is unlikely. However, over time, as AI algorithms become more reliable and trustworthy, perhaps we might see some evolution. But for now, the fundamental principles of data privacy remain paramount.

Q. Data Cloud can ingest unstructured data and create structure around it. Could you elaborate on how this works?

Yes, this is a key capability within Data Cloud. We’ve built a specialized pipeline to ingest various forms of unstructured data like videos, images, PDFs, text documents, chat transcripts, etc. This data is then transformed, chunked, vectorized, and stored in a vector database as meaningful units for optimized search and query needs. 

The beauty of this is that governance can be applied at a granular level over unstructured data. You can classify the chunks of data and then, when queries are performed, our robust support for Retrieval-Augmented Generation (RAG), indexing, and hybrid queries rely on our governance stack to ensure what is being accessed and reasoned over by agents is governed and compliant. 

Q. What role do you see for human oversight and intervention in these data governance frameworks designed for agentic AI? How do we balance autonomy and control?

Human oversight remains critical. We’re actively working on checks and balances to ensure reliable AI models, focusing on preventing bias through diverse training data, rigorous testing, and ensuring algorithmic transparency. 

Human oversight remains critical.

Establishing ethical guidelines and regular audits is also crucial. The interaction between humans and AI agents will be essential for some time, especially in scenarios with strict requirements around bias, explainability, and transparency. We already have infrastructure like chatbots that allows for a fluid transition between agentic and human interaction, ensuring verification and oversight when needed. This hybrid approach will likely satisfy the majority of use cases.

Q. Are there new skills or roles that will be critical for data governance, security, and policy teams to manage the challenges posed by agentic AI?

The majority of the scenarios in the near term will require effective collaboration between autonomous agents and humans. As a result, the ability for an AI agent to operate within a defined scope, knowledge, and guardrails, and then seamlessly hand it off to a human agent when necessary, is key.

Human oversight remains essential, and it is critical that the AI agent and the human agent can both work together effectively across all data types, regardless of where it resides, while having governance oversight around what data they have access to, and what jobs they can perform.

Q. Are customers from larger, more complex industries asking for different things regarding data governance for agentic AI compared to other companies?

Interestingly, ‌fundamental needs are quite similar across different sizes of companies, from SMBs to large enterprises. Everyone is grappling with the implications of AI agents and the associated data, metadata, guardrails, monitoring, and observability. Large enterprises face greater complexity and scale, but the foundational issues we’re addressing with our frameworks are consistent. Ease of use and making these technologies approachable are also universal requirements.

Q. How can organizations establish ethical boundaries to help prevent unintended biases from being amplified as agents parse through data?

This involves ensuring explainability and transparency in AI models, engineering them for debugging, and fostering trust. This principle extends to agents as well. Beyond foundational governance capabilities like a unified policy framework, we’re also focusing on AI-powered tagging to automate metadata tagging with optional human supervision. 

Crucially, we’re building in monitoring and observability capabilities at the governance layer itself. This allows for troubleshooting when policies or classifications lead to unintended outcomes, ensuring the same level of explainability and transparency we strive for in the AI models themselves. We’re working closely with our AI teams to ensure a self-feeding mechanism across our governance and AI testing frameworks.

Q. What else is mission-critical for IT buyers regarding data governance and agentic AI?

The complexity of data fluidity‌ — ‌the fact that data exists in silos‌ — ‌is a key challenge to acknowledge. We need to build systems that can bring action and results to where the data lives and make it accessible to the right roles within the necessary guardrails of trust, scalability, and diverse governance needs. It’s essential to understand your existing ecosystem of external systems and ensure your governance strategy is future-forward. 

We’re actively working on bi-directional integration to allow customers to leverage Data Cloud metadata within our ecosystem together with metadata in third-party data sources that we federate with (zero copy) and other industry leading governance vendors. This metadata interchange allows for consistent application of tags and policies across different governance systems, which is a critical consideration for organizations navigating this evolving landscape.

According to Gartner, organizations will abandon 60% of AI projects that are unsupported by AI-ready data.

According to Gartner, organizations will abandon 60% of AI projects that are unsupported by AI-ready data. In the AI and agent-first enterprise landscape, businesses struggle with fragmented data, complex governance requirements, siloed workflows, limited automation, delayed insights, and misaligned strategies. These challenges hinder the ability to derive meaningful intelligence and deploy trusted AI. In particular, for CIOs, they cite security and data infrastructure as the top reasons hindering AI adoption. Fragmented data governance and complexity remain a key challenge, while data inconsistencies, siloed data sources, scalability challenges, and compliance and security risks remain.

More information:

• Learn more about Salesforce Data Cloud and how it is evolving
• Find more news and stories on data in the agentic AI era, including ​​how Data Cloud powers Agentforce