Marko Tomic

System 02 · AI systems

Multilingual Metadata Automation

3,936 metadata fields across four languages, produced by an agent workflow that earned autonomy - 220+ hours of concentrated manual production avoided.

StatusAutonomous · post-calibration Scope492 entities × 4 languages × 2 fields Review3-4 calibration runs per language

Problem

A four-language catalog with missing metadata

Large portions of the commerce catalog had missing or low-quality meta titles and meta descriptions - across four languages. Writing them manually is skilled, repetitive work: each localized entity requires keyword intent, terminology and tone, not just translation.

The Metadata Automation system operates across the same multilingual commerce catalog described in the International Commerce Transformation case study, and is part of the broader AI Operations architecture.

Scope

What was actually changed

3,936 Metadata fields produced

492 entities × 4 languages = 1,968 localized entities × 2 fields (meta title, meta description).

492 Entities

463 products changed plus 29 collections. Four further products were recognized as already appropriate and deliberately ignored.

220+ h Manual production avoided

Human baseline ~7 min per localized entity → 13,776 min ≈ 229.6 h - roughly 30 uninterrupted 7.7-hour workdays.

Workflow

Review first, autonomy after calibration

  • Autonomous action
  • Human judgment
  • Agent
  • Governance gate
Initial workflow: every field passed through human review in Excel before implementation.
After 3-4 calibration runs per language, output quality held without review - so the human checkpoint was removed. This progression is the point of the system.

Quality rules

What "good" meant

  • Meta Title ≤ 70 characters; Meta Description ≤ 160 characters
  • Duplicate detection across the catalog
  • Product-family keyword alignment
  • Language-specific keyword intent and terminology
  • Calm luxury brand tone - informational, not hyped
  • Avoidance of overly commercial phrasing

Localization is not literal translation: each language's metadata was built around its own search intent and terminology, which is why calibration had to happen per language rather than once.

Outcome & reflection

Autonomy as an earned state

The measured result is 220+ hours of concentrated manual production avoided - not months, not a percentage, just a large block of skilled repetitive work that no longer requires a human to produce it. The more durable result is the pattern: autonomy was granted only after the system demonstrated, per language, that its output survived review unchanged.

Time recovered this way is organizational capacity, not just efficiency - the subject of AI should create organizational leverage.

Related