Google Cloud has provided a concrete figure: its object annotation costs can be 75% lower than those of comparable cloud services, while L'Oréal says its asset-tag processing speed has increased 100-fold. This shows that the bottleneck in enterprise AI is shifting from model capability to data preparation: models can process images and PDFs, but first they need to know what those files contain.
What this is
Object Enrichments, a data-processing capability that automatically converts raw files into data with tags and other contextual information, is a new Google Cloud Storage feature currently in preview. It uses models that also power Google’s internal services to perform object recognition and tagging on images, classify PDFs, and save the results alongside the original files.
It addresses a practical limitation of AI agents: rereading thousands of images, contracts, or video frames for every question consumes substantial computing resources and slows responses. By generating metadata in advance, agents can first narrow down relevant files and then make judgments. RAG (retrieval-augmented generation, which first retrieves relevant content from enterprise materials before passing it to a model to answer) likewise no longer needs to scan all raw materials on every request.
Industry view
Notably, this is not a feature for a general-purpose chat product. It turns “preparing data for agents” into a standard capability within cloud storage. As model capabilities converge, whether data can be understood, invoked, and audited by machines is becoming a practical threshold for enterprise agent deployment. L’Oréal’s claim of a 100-fold increase in tagging speed also suggests that content management and advertising assets may be among the first areas to generate clear value.
But lower costs and better results cannot be equated directly. Google’s claimed 75% savings are based on a comparison with “comparable cloud services,” while actual total cost still depends on file types, model invocation volume, data scale, and downstream retrieval workflows. Pre-extracted tags may also become outdated or contain errors, and once errors are written back alongside the original data, they can create new governance challenges. Saving metadata “in place” makes access easier, but does not automatically resolve permissions, versioning, or data-quality issues.
Impact on regular people
For enterprise IT: Teams may gradually incorporate the organization of unstructured materials into agent projects rather than focusing only on building chat interfaces. At the same time, they will need to add cost accounting, access controls, and manual spot-check mechanisms.
For individual professionals: Searching across documents, images, audio, and video will become more natural-language-driven. In the future, colleagues or business systems may directly invoke these materials to complete tasks, but that does not mean every conclusion produced by a model is reliable.
For the consumer market: Photo organization, product recognition, and media asset management will become cheaper and more automated. Consumers will also encounter more personalized services driven by enterprise materials, at the cost of requiring clearer disclosure of data-use boundaries and error sources.