# Google Cloud Object Enrichments: Scope Your First Bucket Before Scaling

Configure Cloud Storage object enrichments with explicit bucket scope, limited IAM, eligibility checks and a cost review before processing an archive.

By Paras Madan | 2026-11-19


A searchable archive needs more than files in a bucket. It needs useful descriptions, a way to find missing descriptions, and a clear answer to who can generate them.

Google Cloud introduced object enrichments in preview on November 19. The service generates metadata for stored images and PDFs and attaches it to the objects themselves. The practical starting point is not your entire archive. It is one bounded collection with a retrieval problem you can measure.

## Separate the preview from the roadmap

The launch covers five models: object detection, image labeling, face detection, explicit-content detection and PDF classification. Gemini-based summarization, transcription and other complex multimodal enrichment are described as upcoming functionality requiring an access request, not something to assume is enabled with this preview. [Google's announcement](https://cloud.google.com/blog/products/storage-data-transfer/cost-effective-ai-context-extraction-from-unstructured-gcs-data) makes that distinction.

For a proposed invoice-search pilot, start with PDF classification. Do not promise that a category label extracts invoice numbers, validates totals or resolves duplicate vendors. Those are separate requirements to evaluate before changing the accounting workflow.

Write the pilot question narrowly: can the team find the right document class with fewer irrelevant results? That gives you an acceptance criterion instead of a collection of impressive-looking annotations.

## Check eligibility before interpreting missing results

The [object enrichments overview](https://docs.cloud.google.com/storage/docs/object-enrichments/overview) sets several important boundaries. Objects must be no larger than 20 MiB and carry a supported Content-Type. A valid PDF with incorrect or missing metadata can be skipped. Rapid buckets, hierarchical-namespace buckets and objects encrypted with customer-supplied keys are unsupported.

Processing locations are `us`, `eu` and `asia`. Configuring the same bucket and model in multiple locations causes processing and billing in each location. Configuration changes can take up to 24 hours to propagate, while deletion typically takes a few minutes.

Build an eligibility ledger before enabling the pilot: object identifier, size, content type, expected category and inclusion decision. Keep skipped files in that ledger. Otherwise, a clean-looking result set can conceal the documents the system never evaluated.

Use deliberately awkward examples as well as clean ones: an invoice saved with a generic content type, a large PDF and a document that a human would classify differently depending on context. 

## Scope configuration and permissions separately

A FeatureConfig selects buckets, models and processing location. Bucket names must be exact, despite the API field name suggesting regex support. Newly created or recreated buckets are not automatically included in an existing configuration.

The console can grant project-wide service-agent permissions. For a narrower pilot, the [configuration guide](https://docs.cloud.google.com/storage/docs/object-enrichments/configure-object-enrichments) also supports granting `roles/storage.annotationGeneratorService` on individual buckets through the CLI, with `roles/storage.bucketViewer` at project level. Creating a configuration alone does not authorize processing.

I would use a dedicated pilot bucket and record two independent approvals: which objects may be processed, and which identity may access them. A narrow configuration is not a substitute for narrow permissions, or vice versa.

Choose one owner for both records. If another team adds a model or expands bucket scope, that owner should review the change before it becomes the new baseline. 

## Measure the bill around the model

Inference is only one cost component. Google documents metadata storage, data access operations and networking charges; Storage Insights and BigQuery analysis can add costs. Reads can also trigger Autoclass storage-class transitions.

For the invoice pilot, use a simple internal worksheet rather than extrapolating from a promotional per-object figure. Record the number of eligible documents, selected models, processing locations, observed inference charges, storage and access charges, and any downstream querying costs.

Then compare cost per useful retrieval outcome, not just cost per processed file. A cheap label that sends a reviewer to the wrong document can be more expensive operationally than leaving the file unclassified.

Keep query experiments separate from enrichment experiments in that worksheet. Otherwise, a busy analyst exploring results could make the enrichment service appear more expensive than it was, while an unused metadata store could make it appear more valuable than it is.

## Review classifications before routing work

Sample successful classifications and missing results separately. For each returned document, ask whether its category helped retrieval, whether the category was misleading, and whether a reviewer needed to reopen the original anyway.

A useful pilot report has three groups: helpful, wrong and unresolved. Do not force the unresolved group into success or failure merely to simplify a chart. 

Keep downstream actions reversible during the pilot. A label can suggest a review queue without becoming permission to approve a payment or delete a file.

## Plan the stop condition

Deleting a FeatureConfig stops new enrichment but leaves existing annotations attached. Treat stopping processing and cleaning up metadata as separate decisions.

Before scaling, name the person who can stop the pilot, the budget threshold that triggers review and the quality failure that blocks expansion. Expand only when the same configuration can be explained alongside its costs and retrieval results.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
