Case study · May 2026
AI Metadata Enrichment Pipeline
How to operationalize AI at enterprise scale without sacrificing reliability, governance, or trust.
OutcomeMillions of assets enriched in production, each in seconds to a minute, and in its first year the work of 550+ skilled professionals.
Manual metadata creation does not scale. As asset volumes reach the millions, metadata becomes inconsistent, incomplete, and expensive to maintain, and the gaps compound downstream: search quality degrades, workflows slow, and the organization loses visibility into how assets are used. Automation is the obvious answer. Introducing AI into a production environment without guardrails creates a different set of problems, including unreliable output, errors that are difficult to trace, and rework at a scale that erases the savings. The pipeline had to address both.
The pipeline runs inside the platform's production environment alongside the asset workflows the organization depends on, so reliability and performance were requirements rather than goals. It also had to satisfy partners in legal and content strategy who needed to know what generated a given piece of metadata and when.
The central design decision was to treat enrichment as a governed process rather than an autonomous one. Only approved prompts and model configurations run, and every enrichment record traces back to the configuration in use at the time. Enrichment can be re-run when inputs change, when models improve, or when a problem is found, without reconstructing what happened. The system favors repeatability and observability over novelty, which is the correct tradeoff for a pipeline that runs continuously.
- INPUTAsset
- INPUTApproved prompt
- INPUTModel configuration
- Execution record
- Configuration version
- Timestamp
The pipeline does not test new prompts. It executes approved ones. Evaluation happens in a separate environment, the Prompt Playground and Bulk Testing Suite, and a configuration reaches the pipeline only after it has met the criteria agreed there. Promotion from evaluation to production is an explicit step with a record, which keeps untested behavior out of live workflows while evaluation continues at its own pace.
The pipeline processes millions of assets at a scale manual effort could not approach, and has made metadata consistent enough that downstream search and automation can depend on it. It also established a repeatable pattern for adding AI to production systems: intelligence run as governed infrastructure, with traceability designed in from the start rather than added after a problem.