VaikoraVaikora

Vaikora › Blog › Product Updates

AI agents and your data governance program: where the enforcement gaps actually appear

Product Updates · Nov 16, 2026 · 4 min read

Every data governance program has a set of controls: classification policies, retention schedules, access rules, audit logs. Those controls were designed for humans and applications that behave predictably. AI agents do not behave predictably, and most governance programs have not caught up.

The gap is not theoretical. An AI agent authorized to read customer records can, in a single session, extract, summarize, reformat, and transmit data across system boundaries. Each individual step may look permitted. The aggregate action violates your data governance policy. By the time an audit surfaces it, the data has moved.

Where traditional data governance stops

Data governance frameworks typically enforce at four points: classification at ingestion, access control at the data layer, retention at the storage layer, and audit at the query layer.

None of those controls sit between an AI agent and the action it is about to take. A role-based access control policy says whether the agent's service account can read a table. It says nothing about whether the agent should be allowed to send that table's contents to an external API, include it in a generated document, or pass it as context to another model.

The enforcement gap is at the action layer, not the data layer.

What an action-layer control looks like

Vaikora sits inline between an AI agent and every LLM call or tool invocation it makes. Before the action executes, Vaikora evaluates it against deterministic policy and returns one of four decisions: allow, block, modify, or require approval. Identical input always produces the same decision. There is no probabilistic scoring that shifts based on model temperature or prompt phrasing.

A concrete example: an agent is processing a support ticket that contains a customer's payment card number. The agent constructs a prompt that includes the raw PAN to give the model context. Vaikora inspects the prompt before it reaches the model, detects the PAN pattern, and either redacts it or blocks the call entirely, depending on how the PCI-DSS compliance preset is configured. The customer record never reaches the model in unmasked form. The decision is written to a SHA-256 hash-chained audit log.

That audit log entry is not a flag for review. It is a tamper-evident record of what the agent attempted, what policy applied, and what happened, in order, with cryptographic integrity.

How this connects to your existing governance program

Governance programs already have classification schemes. The enforcement question is whether those classifications travel with the data when an AI agent touches it.

Vaikora policy can be written against data classifications directly. If a field is tagged as containing PII or PHI, policy can require that any agent action involving that field either redact the sensitive value, require human approval before the action runs, or block the action outright. The classification work your team has already done becomes an input to runtime enforcement, not just a metadata label in a catalog.

Vaikora also adds a runtime gate on top of existing controls in Azure RBAC, AWS IAM, and Snowflake roles rather than replacing them. Your access control model stays intact. What you gain is a control layer that evaluates intent and content at the moment of execution, which IAM cannot do.

Compliance preset coverage

For teams operating under specific regulatory frameworks, Vaikora ships presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. Each preset defines a baseline set of policy rules tuned to the data handling requirements of that framework. You can extend or restrict the preset to match your internal policies.

The audit output from Vaikora supports compliance review directly. Because every decision is hash-chained and tamper-evident, the log can demonstrate to an auditor that a specific agent action was evaluated, that a specific policy applied, and what the outcome was. That is a different category of evidence from a system log that records only what did execute.

What to check in your current environment

Start with two questions. First, which AI agents in your environment have read access to data your governance program classifies as sensitive? Second, what controls evaluate those agents' actions at the point of execution, before data leaves the system?

If the answer to the second question is "none" or "only after the fact," the action layer is uncontrolled. Map your highest-risk agent workflows, identify the data classifications those agents can reach, and determine whether your current controls can enforce policy at the moment an action runs. If they cannot, that is the gap to close.

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

More from the Vaikora blog