AI-driven business requires a new level of speed and poses whole new risks. Network-level access and security won’t cut it anymore. It’s time to bring access and governance to the data directly.
Artificial intelligence has fundamentally changed the relationship between software and data. For the first time, systems don’t just query information—they reason over it, act on it, and chain decisions together at machine speed. And even more challenging from a security standpoint, AI delivers more value as it brings in more data – which means security-minded firms feel forced to walk a tightrope between providing ample training data while also keeping information close to the vest.
Yet most enterprise data architectures remain anchored to assumptions from a pre-AI era: static workloads, predictable access patterns, widespread data duplication as the default, governance applied after the fact, and building apps for pre-defined needs and roles — not the constantly-evolving directives of agents and LLMs.
That mismatch now poses a major threat to AI-era security — and looms as a primary blocker to meaningful AI adoption. Enterprises are discovering that deploying AI safely and at scale isn’t a model problem or a tooling problem — it’s a data architecture problem. To be AI-ready, enterprise data must meet a new set of requirements that many existing systems were never designed to support. They must move data access — and data security — to the data plane.
Data Access Must Be Immediate
In an era of autonomous, in-the-moment decision-making, there’s no time to wait for latency. You need the data now. For example:
- When your factory’s agentic Operational Technology (OT) needs sensor data to coordinate a shipment, your existing stack may not have the necessary real-time data at the ready – and currently available data lake / data warehouse information may already be stale.
- If you’re a large telecom firm using AI assistants to chat with customers, you’ll need instant access to customer profiles, past tickets, billing info, service outages, and policy rules to make sure every conversation is relevant and informed. If data is slow, fragmented, or batch-based, your AI will give generic or outdated answers – and a bad customer experience. Complicating matters dramatically, all that fast access must happen with privacy at the center – ensuring that, in the rush to access pertinent data quickly, no sensitive customer information gets exposed in the process.
- If you’re a credit card company using AI models to detect fraud, you’ll need immediate access to transaction streams, historical spending patterns, device data, and geolocation signals. That’s because many fraud identification decisions happen at the moment of transaction – and need to operate in milliseconds, not minutes.
The examples are nearly endless, but the underlying logic is the same. AI-era decisionmaking requires near-instant information, insight and just-in-time policy enforcement based on an employee’s permissions. Waiting on the data as it winds through intermediaries, networks, and “shipments” won’t cut it. And to make that data accessible, companies must shift the path to data completely. Firms must take their AI applications to the relevant data directly.
Make “In Place” the New Default
Copying and sharing data has always been a risky proposition. Whether you’re passing information along to an external app, standing up a staging environment or even sharing across teams internally, every data motion presents another surface to attack — and another potential loss of fidelity.
But while just a few years ago those hazards could present a justifiable risk/benefit trade-off, data movement has become an untenable liability in the age of AI. Agentic threats can be more sophisticated, harder to observe and control, and faster-acting than traditional threat sources. Meanwhile, even accidental, non-malicious data leakage can turn your sensitive data and IP into a foundational training source for someone else’s AI.
These are rarely bearable risks, no matter what the upside. For this reason, AI-era data needs to stay fully protected in exactly one state: in place.
The two changes above are dramatic — and call for a radical shift in how we govern, access, and manage data for agents and LLMs across the enterprise. We’ll need to meet the four new requirements — outlined below — to make this shift possible.
Requirement #1: Go Straight to the Data
The only way to ensure truly real-time access to the data is to enable real-time access to the original data, where the data resides. This means no more waiting on information to flow through networks (and waiting for manual sends and approvals) to get the immediate information that AI needs.
With the right governance in play (see Requirement #2), this also means a marked security upgrade over traditional computing. If firms can go straight to the data to support AI, they don’t need to copy and share data across the org – and they can significantly decrease attack surfaces and potential points of data leakage as a result. Going straight to the source – not to a copy – is obviously a win for data fidelity as well.
Requirement #2: Go Beyond the Network. Protect the Data
Achieving the above requires a complete shift in data access — which in turn takes a sea-change in data governance. Rather than protecting a network to secure the data roaming freely behind its walls, firms must control the data directly. The architecture of data protection must shift from sweeping governance across teams and whole firms, to targeted governance around the specific data object at the moment of use, before data ever touches a model, agent, or automated workflow.
Requirement #3: Metadata Becomes the Control Plane
AI systems don’t need raw data to reason—they need structure, context, and rules. That makes metadata the natural control plane for AI-era governance.
This is a huge asset that can be used in AI data security. Rather than granting agents or applications direct access to sensitive tables or files, enterprises should require AI workflows to operate on metadata alone: schemas, semantics, sensitivity classifications, lineage, and policy constraints. Models can then generate execution plans or “recipes” that are safely evaluated and executed under strict controls.
This approach dramatically reduces exposure while preserving flexibility, and it enables AI systems to operate safely across regulated and mission-critical environments.
Requirement #4: Execution Must Be Isolated and Auditable by Default
AI-generated code and queries cannot be treated like human-authored requests. They must execute in isolated, hardened environments with narrowly scoped permissions and full auditability.
Enterprise AI readiness requires:
- Isolated execution sandboxes
- Ephemeral, least-privilege credentials
- Policy-based enforcement at every step
- Field-level logging and traceability
- Explicit prevention of unauthorized joins, exports, or data movement
Without these controls, AI systems introduce silent risk that compounds over time.
Requirement #5: AI Access Must Work Across Standards and Environments
Enterprises don’t operate in a single cloud, toolchain, or protocol—and neither does AI. AI-ready data architectures must support multiple interfaces and standards, from APIs to agent frameworks, without fragmenting governance or security models.
This requires tight integration across access, policy, transformation, and execution layers, so AI systems can evolve without forcing enterprises to re-architect security every time a new framework or standard emerges.
The Shift Enterprises Must Make
AI readiness is not about adopting more tools or adding more guardrails. It’s about redefining how data is accessed, governed, and executed in environments where machines—not humans—are the primary consumers.
Enterprises that succeed will be those that treat governance as a prerequisite, not an afterthought; that keep data in place rather than moving it; and that elevate metadata, policy, and auditability to first-class system components.
This architectural shift is already underway. The question for most organizations is no longer whether their data strategy must change—but how quickly they can make it AI-ready without compromising security, compliance, or trust.