
Cloud Storage object enrichments: AI context for unstructured data

Enterprise data is mostly unstructured, and raw objects in Cloud Storage are meaningless bytes to AI agents. Google Cloud is introducing object enrichments, a managed Cloud Storage capability in preview that automatically extracts rich contextual metadata from objects and saves it in place as object context, so agents can reason over grounded knowledge without reading every file at runtime. The service provides access to AI models used by internal Google services like Google Photos, including object detection, face detection, image labeling, and explicit content detection, at what Google states is 75% cost savings compared with similar services from other cloud providers. It also includes a PDF classification model that labels documents into categories such as Finance, Games, News, and Health.
The announcement frames this as a fix for three historical problems in preparing unstructured data for agents. First, running models over millions or billions of objects is expensive, leading teams to process only a fraction of their data and leaving blind spots. Second, self-managed pipelines require stitching together event triggers, queues, custom infrastructure on Cloud Run or GKE, and scalable retry/error logic, with continuous upkeep. Third, extracted metadata often ends up in a separate store, decoupled from the original bytes, forcing duplicate storage, separate permission models, and cleanup jobs to keep metadata in sync.
The headline cost number is 10 million images or PDFs enriched for $700 according to published pricing. In addition to first-party image models, object enrichments will natively support Gemini for multimodal reasoning over text, PDFs, images, audio, and video, enabling use cases like summarization, entity extraction, transcription, and translation; that functionality is not yet generally available and requires contacting Google. Hyperscience reports using object enrichments to drive a 50%+ lift in document classification accuracy while reducing classification inference costs by 2-3x.
Operationally, setup is described as three no-code steps in the console or via the Cloud Storage API: define the bucket scope and processing location, choose one of the five first-party models, and grant IAM permissions. The service then orchestrates model inference, parses responses, writes results into each object’s object context, and handles retries and throttling. Because object context is a Cloud Storage metadata construct stored alongside the object, it shares the same durability and consistency properties, travels with the object during copy, move, and replication, and is automatically deleted with the object. IAM controls read/write/update/remove access to context, and applications can filter object listings by context key-value pairs. Context can also flow natively into BigQuery via Storage Insights datasets for SQL analytics without ETL code.
Early customers are using object enrichments for creative asset search and media archives, automated trust and safety moderation, training data curation for autonomous driving and robotics, and multimodal analytics. L’Oreal says object enrichments lets it label its creative asset repository 100x faster at a fraction of the cost of existing pipelines, and Vivint reports that object enrichments combined with BigQuery cut AI development from 6 months to 1 month. For use cases like finding ‘a red car driving at sunset’ or moderating millions of user uploads, the enriched labels can be indexed into vector databases or BigQuery to power RAG and asset-discovery agents.
Object enrichments is available in preview today through the Cloud Storage console and Cloud Storage API. Google points to documentation, BigQuery setup, pairing with the Cloud Storage MCP server for IAM-scoped agent access, and published pricing as next steps, with Gemini enrichments available on request.

