Vaikora › Blog › Product Updates
Classification at runtime: how Vaikora enforces data policy on what an AI agent actually touches
A file without a classification label is a file your controls cannot act on. The label is not metadata decoration. It is the signal that tells every downstream gate whether to allow, block, or escalate. Without it, enforcement is guesswork.
This matters especially when AI agents are in the loop. An agent reading a CRM export, pulling from a data warehouse, or writing to a document store does not pause to ask whether the data is sensitive. It executes. If your classification exists only in a governance dashboard and not at the point of action, the agent never sees it.
What "classification at the source" actually means
Classification at the source means the label travels with the data from the moment it is created or ingested, not applied retroactively by a scanning job that runs weekly. A document tagged CONFIDENTIAL in SharePoint when it is saved carries that label into every system that reads it. A database column marked PII at the schema level surfaces that attribute when queried.
The practical difference is timing. Retroactive classification finds sensitive data after it has already moved. Source classification gives enforcement systems something to act on before the data moves.
The two mechanisms that make this work in practice:
- Persistent labels attached to files, records, or fields at creation, using formats that survive export, copy, and format conversion
- Schema-level metadata on databases and APIs that tags columns or response fields as regulated, so any consumer inherits the classification context
When both are present, a downstream control knows what it is handling without re-scanning.
Where the control gap opens in AI agent workflows
Consider a data pipeline where an AI agent is authorized to query a Snowflake table that contains customer records. The agent has a valid service account, Snowflake row-level security permits the query, and the agent returns results to a downstream process.
If that table has no classification on the columns it returns, the agent's tool call looks like any other permitted query. There is no signal for a policy engine to act on. The data leaves the warehouse, enters the agent context window, and may be passed to an LLM or written to an output file. At no point did a control see "PII" or "PCI" and make a decision.
Now apply source classification. The column is marked PII at the schema. When the agent calls the tool, the policy engine reads the classification metadata before returning results. It can enforce: allow the query but strip the column, require approval before the full record is returned, or block entirely based on the destination.
The control is not invented at the enforcement layer. It is inherited from the classification that was already there.
How Vaikora uses classification as an enforcement signal
Vaikora sits inline between an AI agent and every LLM call or tool execution. Before an action runs, Vaikora evaluates it against deterministic policy. Classification labels are a first-order input to that policy evaluation.
When a tool call carries data marked with a sensitivity label, or when Vaikora's content inspection detects regulated content in the payload, the enforcement decision is based on what the data actually is, not on the identity of the system it came from. That distinction matters. A service account with broad IAM permissions is not the same as a service account that is explicitly authorized to handle PII in this context.
Vaikora ships compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. Each preset defines which classification labels or content patterns trigger which enforcement outcomes. The preset does not replace your classification scheme. It reads it.
Every enforcement decision is written to a SHA-256 hash-chained audit log. If a regulator or internal audit team asks what happened to a specific record, the log shows the classification that was present, the policy that applied, and the outcome, in order, tamper-evident.
Vaikora also adds a runtime gate on top of existing controls like Azure RBAC, AWS IAM, and Snowflake roles rather than replacing them. Classification flows through those layers and into Vaikora's inline enforcement point.
What to check in your current environment
Before adding more enforcement tooling, audit what you actually have:
- Do your database schemas carry column-level sensitivity metadata, or is classification only in a separate catalog?
- Do files exported from regulated systems retain their labels when opened in downstream tools?
- When an AI agent queries an API, does the response include classification context, or does that metadata stay behind in the source system?
If classification exists in a governance tool but does not reach the systems where agents operate, the enforcement you think you have is not enforced at the point that matters.
The place to start is the gap between where classification is recorded and where action is taken. Close that gap first, then configure enforcement policies that read the labels that are actually present.
See Vaikora enforce policy on your AI
Open-core AI runtime control. Self-host the MIT gateway free, or run the hosted Control Plane.
Get a demo Self-host the gateway
Vaikora