Enterprises across industries are generating unstructured data at a staggering pace—scanned contracts, design blueprints, call-center audio, clinical imaging, compliance documents, product manuals, surveillance video, and high-resolution media assets. For many organizations, this category represents the majority of new data generated each year, yet it remains the least accessible. Senior data and technology leaders frequently surface a common challenge in strategic discussions: How do we manage, govern, and extract value from unstructured data when we don’t even know what is inside these files? This question appears repeatedly in customer conversations, and several enterprises have already taken guidance from a reference architecture designed to address it.
The core difficulty begins with cataloging. While structured datasets come with schemas, columns, and well-understood metadata, unstructured files—especially media formats—arrive as opaque binaries. A storage system has no inherent ability to understand whether a document contains sensitive customer identifiers, whether a video shows a safety compliance incident, or whether a scanned PDF includes handwritten notes that alter legal interpretation. Without visibility into the contents, it is impossible to build an effective data catalog, enforce retention rules, or apply access controls. The lack of semantic context prevents classification, analytics, lineage mapping, and policy enforcement. In effect, organizations are storing massive volumes of data that they can neither govern nor use confidently.
To solve this, Sakti Mishra, Principal Solutions Architect at AWS, introduced a cloud-native architecture that applies AI-driven enrichment as a prerequisite step before cataloging. The approach was published in the AWS Big Data Blog titled “Unstructured data management and governance using AWS AI/ML and analytics services.” The design begins by storing media and document files in their raw form on Amazon S3, preserving original content as the source of truth. Instead of attempting to define metadata upfront, the system triggers AI workflows after ingestion. Services such as Amazon Textract, Amazon Comprehend, Amazon Rekognition, and Amazon Transcribe extract interpretable signals from documents, images, and audio, while custom models on Amazon SageMaker and foundation models on Amazon Bedrock add domain-specific interpretation. These outputs are stored as structured data—JSON artifacts, tabular metadata, embeddings, or processed derivatives—linked back to the original assets.
This sequencing is a key innovation: enrichment precedes cataloging. Once structured metadata exists, organizations can finally apply mature data-lake governance patterns—business glossaries, search layers, lineage graphs, and query engines—just as they would for relational datasets. The catalog becomes dynamic, capable of evolving as new AI models extract deeper or more accurate information, ensuring historical content remains useful without re-ingesting or retrofitting storage systems.
Governance is equally central to the design. Instead of allowing users direct access to raw content, the architecture introduces a layered interaction model. Users begin by exploring the catalog, which provides descriptive context without revealing sensitive underlying assets. If permitted, they can interact with AI-extracted data, gaining analytical insight without direct exposure to raw files. Access to original media objects is granted only when required by role or entitlement, enforced through fine-grained access controls using AWS services such as Lake Formation, DataZone, and S3 Access Points. This structure enables secure collaboration across compliance teams, analysts, machine learning practitioners, and business units, allowing each to work at the appropriate level of abstraction.
In the heart of this reference architecture, it proposes a refined, three-tier system that elegantly bridges raw unstructured content, AI-driven semantic enrichment, and governed metadata discovery. As detailed in his post, this design underpins a practical, scalable, and secure way to manage unstructured data at enterprise scale.
High-Level Architecture: A Layered Flow from Raw Data to Governed Insight
At a high level, the architecture diagram Sakti Mishra and his colleagues at AWS, presents a tiered flow that carefully separates raw content, AI-extracted output, and a metadata catalog.
The diagram represents the following layers.
- Raw Ingestion Layer
The process begins with raw unstructured objects—images, video files, scanned documents, audio, and more—landing in an S3 bucket (often called the “raw input object store”). These files are preserved in their native form, providing a source-of-truth repository that retains full fidelity.
- AI Enrichment Layer
Once ingested, the architecture invokes multiple AI and ML services to analyze the content. Depending on the file type, services like Amazon Textract (for scanned documents), Amazon Comprehend (for natural language), Amazon Rekognition (for images and video), and Amazon Transcribe (for audio) are used. Custom or domain-specific models running on Amazon SageMaker or via foundation models from Amazon Bedrock can also be integrated to extract richer metadata or embeddings. The output—structured JSON, embeddings, labels, and other ML artifacts—is written back to S3 in a separate “extracted JSON” bucket.
- Metadata / Catalog Layer
Once enriched, a metadata layer is created to maintain relationships between the raw objects and their derived interpretations. This layer stores mappings between original S3 paths, the enriched JSON or structured outputs, and any additional enrichment transformations. According to the diagram, this metadata layer is the foundation for discovery: it enables cataloging systems to expose what is inside media assets without compromising the raw content.
- Query and Governance Layer
On top of the metadata catalog, flexible query engines are leveraged for discovery and analysis. The architecture calls out Amazon Athena and Amazon Redshift Spectrum to query the structured AI outputs. The AWS Glue Data Catalog is used to register and organize schema metadata. Governance is enforced at multiple levels: AWS Lake Formation and Amazon DataZone manage fine-grained access controls to both the extracted data and the raw objects.
- Controlled Access to Raw Assets
Finally, when users actually need to access the original media—for example, a compliance team needing to review a video file or legal requiring a scanned contract—the system uses Amazon S3 Access Points. These access points enforce restrictive policies, controlling access by prefix or tag, ensuring only appropriately privileged roles can retrieve the raw binaries.
Why This Architecture Stands Out
This design offers the following two unique approaches that addresses the complex requirement around unstructured data management.
- AI-First Metadata Strategy: Rather than cataloging based on limited or manually assigned tags, the architecture leverages AI to discover what’s inside It effectively builds a structured layer on top of unstructured content. Once in structured form, traditional cataloging and governance patterns become applicable.
- Multi-Level Access Control: By separating metadata, AI outputs, and raw assets, the architecture enables granular access. Users gain insight and perform analysis without automatically gaining access to potentially sensitive raw files, unless explicitly authorized. This staged flow provides powerful control, auditability, and security.
Importantly, the architecture is technology-agnostic. Although the AWS blog demonstrated it using AWS-native services, the same pattern can be implemented using other cloud or on-prem platforms, AI frameworks, storage systems, and metadata/catalog tools. As long as a system can capture raw data, perform enrichment, store derived artifacts, and maintain a catalog, it fits within this paradigm.
Early adopters of this approach have already applied the model to real-world domains such as regulatory document management, digital media archives, contact-center analytics, and healthcare information retrieval. They view this architecture not as a storage strategy, but as a knowledge-extraction framework—one that retains the fidelity of original files while enabling structured insight and controlled access. It creates a continuity between traditional relational governance and the rapidly expanding world of multimedia-driven datasets.
The model naturally extends into semantic search, where vector embeddings produced during enrichment allow indexing based on meaning rather than keywords. By integrating services such as Amazon OpenSearch Service or open-source Elasticsearch, enterprises can retrieve similar documents, correlate investigative media assets, or identify patterns across diverse formats—even when terminology differs across files. This shifts the enterprise from file storage to contextual knowledge retrieval, creating new value streams across operations, risk assessment, and innovation pipelines.
Ultimately, this pattern reframes the industry’s approach to unstructured data. Rather than forcing media files into rigid schemas or leaving them idle in object stores, the architecture layers AI-driven understanding, structured cataloging, and policy-aligned access control in a cohesive system. By enabling unstructured datasets to behave like governed analytical assets, organizations can finally unlock value that was previously invisible—turning raw archives into actionable intelligence.
By applying this layered, AI-led, catalog-driven approach, organizations can finally address the common-but-difficult questions: What unstructured content do we have? What’s inside those files? Who should see them? And how do we make sense of them without sacrificing security or scalability?
