Enterprises store petabytes of unstructured images, PDFs, audio, and video in object storage, but AI agents cannot reason over raw bytes directly, according to Google Cloud's announcement. Streaming thousands of camera frames into a model on every user prompt burns expensive tokens and adds minutes of delay, the blog post says. Building self-managed preprocessing pipelines with event triggers, queues, and Cloud Run or Google Kubernetes Engine often takes months of engineering toil. Storing extracted metadata in a separate database also forces teams to govern disconnected permissions and run custom cleanup jobs when source files move or are deleted, Google Cloud says.1

Object enrichments replaces those custom pipelines with a managed workflow inside Cloud Storage, Google Cloud says. Builders define target buckets and a processing location, select from five first-party models for images and PDFs, and grant IAM permissions through the console or the Cloud Storage API. The service then orchestrates inference, parses responses, and writes results directly to each file's object context with built-in retries and throttling, according to the post. Google Cloud says the service scales to millions or billions of objects without queues or glue code to maintain.1

The preview gives customers access to internal models used by services like Google Photos, according to Google Cloud. For images, the service supports object detection and image labeling alongside face and explicit content detection, the blog post says. Its PDF classification model categorizes documents into standard industry categories such as Finance and Health. Google Cloud says the service costs $700 to process 10 million images or PDFs and delivers 75% cost savings on image annotations compared to similar cloud services.1

Early customer deployments show speed and accuracy gains over existing pipelines, according to quotes in Google Cloud's post. A principal data and AI architect at L'Oreal says object enrichments lets the company label its creative asset repository 100x faster at a fraction of prior pipeline costs. A vice president of partnerships at Hyperscience says the service drove a 50%+ lift in document classification accuracy while cutting classification inference costs by 2-3x. Neither customer quote specifies the dataset size or test methodology used to produce those figures.1

The service is currently in preview, and its broader multimodal capabilities remain gated, Google Cloud says. Native Gemini integration for tasks like summarization, entity extraction, transcription, and translation across audio and video is listed as upcoming functionality. Teams that want to use Gemini with object enrichments must email Google Cloud to request access, according to the announcement. The post does not state when the preview will reach general availability or disclose pricing for the upcoming Gemini tier.1