Bedrock Data announced Bedrock Data ArgusAI and Natural Language Policy. ArgusAI is a new product that expands the company’s capabilities into artificial intelligence governance. It allows enterprises to understand what data their AI models and agents access during training and inference, and evaluates whether existing guardrails prevent sensitive data leakage.
Natural Language Policy enables teams to implement data controls across all systems using plain English. When combined with ArgusAI, teams can also govern their AI systems using natural language.
Existing tools cannot validate guardrails or enforce policies at enterprise data scales. ArgusAI closes this gap by leveraging Bedrock’s petabyte-scale Metadata Lake, enabling organizations to govern AI with the same context, precision, and automation already applied to structured, unstructured, and SaaS data across cloud and on-premises environments.
Closing the governance gap between data and AI
As organizations build on services such as Amazon Bedrock, new risks arise from opaque data usage and inconsistent controls. DSPM tools stop at discovery, they cannot explain what information trains a model, what AI can see during inference or whether guardrails truly prevent sensitive data from being exposed.
Bedrock Data ArgusAI addresses these challenges through two core capabilities:
- AI Data Bill of Materials (DBOM): Automatically links Amazon Bedrock Custom Models, Knowledge Bases for retrieval-augmented generation (RAG) and Agents to the datasets used for training and inference.
- Guardrail gap analysis and remediation: Evaluates whether technical guardrails are sufficient to block sensitive data exposure based on the DBOM. When gaps are found, ArgusAI generates targeted remediations that security, product and ML engineering teams can apply directly.
Together, these capabilities ensure that AI systems respect enterprise data policies by design, not by manual oversight.
Guardrail gap analysis identifies insufficient controls and generates precise remediations
ArgusAI performs Guardrail Gap Analysis by leveraging Bedrock Data’s Metadata Lake, a unified graph knowledge base that maps the full context of enterprise data including sensitivity, lineage, ownership and entitlements across IaaS, PaaS, SaaS and on-premises systems. Because the Metadata Lake can uniquely scale to enterprise data at petabyte scale, the DBOM generated for AI models provides the most complete record of model-data relationships available.
Guardrail Gap Analysis draws directly from this foundation, comparing policy intent against model exposure to suggest targeted control updates. For example, if policy states that agents must not process or reveal credit-card information, the Bedrock Data platform identifies that such data appears in a model’s training or inference path, determines that existing guardrails are insufficient and produces a concrete remediation plan to suppress or mask the data at the source or agent runtime.
Natural Language Policy and Investigation provides a unified governance layer across data, identity and AI
Natural Language Policy and Investigation enables legal, GRC and security teams to articulate policies and perform investigations in plain English. When applied alongside ArgusAI, natural language governance becomes a platform-wide capability. Teams can articulate policies in plain English or provide policy documents from legal or governance teams, such as “Only members of HR may view employment data” or “EU personal data must remain within EU regions.” Bedrock Data translates these invariants into enforceable controls across connected systems.
This capability extends seamlessly over the DBOM and guardrail context for AI, allowing investigators to ask questions like “Which models can access PHI data?” or “Which identities interacted with customer PII through an agent?” The system responds with precise, explainable answers derived from Bedrock Data’s Metadata Lake, complete with lineage and access evidence, uniting data, identity and AI governance under one operational language.
“Enterprises face a dual challenge: accelerating AI adoption while ensuring governance keeps pace. The fundamental problem is that policies can’t be enforced across fragmented systems that each speak different control languages,” said Pranava Adduri, CTO of Bedrock Data. “When business teams ship AI quickly and governance can’t keep up due to system complexity, the gap between policy intent and enforcement becomes unacceptable. By tying every model to its DBOM, assessing and fixing guardrail deficiencies and translating natural language into enforceable controls, ArgusAI makes responsible AI operational at scale.”