5 minutes
h1
style
article-header-section

Inconsistent manual tagging and unstructured AI tagging can undermine asset discoverability, but pairing manual and AI tagging with human oversight can help teams achieve scalable, high-quality metadata.

Manual tagging often creates inconsistent metadata that can hinder searchability, asset retrieval, and personalization. While AI is a powerful accelerator for context-driven tagging, it amplifies bad data without a solid foundation. To scale, organizations should utilize a hybrid governance model that balances AI automation with human-driven confidence thresholds, approval workflows, and iterative auditing.

As organizations scale their digital footprints, the volume of assets created and promoted is unmatched. These assets need to be housed in a central location making it easier for Marketers to store, organize, search and find relevant assets. When assets are stored in a Digital Asset Management (DAM) tool like the Adobe Experience Manager (AEM) DAM, the metadata associated with each asset is pivotal to how the system will deliver assets to end users via relevant search results. With the high volume of assets residing in a DAM, manual tagging can become a bottleneck, slowing asset discoverability and use. Luckily AI can help!

Smart tagging, or AI-tagging, can help Librarians and Marketers alike apply tags automatically and ease associated governance processes; however, while it is tempting to view AI as a magic wand that instantly organizes a chaotic system, the reality is more complex.

AI does not replace taxonomy or human governance but rather it amplifies bad metadata if your foundation is weak.

The role of the DAM team is actively shifting from manual data entry to AI system auditing. AI might be the processing power, but the DAM team writes the source code, the controlled vocabularies and foundational metadata that makes it run correctly. After all, AI is only as smart as the data we give it.

The hidden cost of human tagging

When tagging relies entirely on human interpretation, consistency becomes a casualty of perspective. Even with dedicated librarians and authors, a single ambiguous tag can break an entire metadata system.

Let’s explore a use case centered around a fictitious event called “The Horizon Summit.” For the exact same asset or page, different teams can and will interpret its context differently:

Each interpretation is valid in isolation. Together, they create chaos.

This inconsistency leads to broken search results and negatively impacts personalization. If a user searches for the 2026 event, they might get served 2025 content simply because the tags lack strict contextual boundaries leading to poor search results, user experience, and ultimately wasted time.

The core challenge isn't human error. It is that humans interpret context differently, and metadata systems were never designed to accommodate that variability without structure.

Understanding your AEM AI tagging options

Before deciding on a governance model, it's important to understand that AEM Assets offers two distinct AI-powered tagging capabilities—each with different strengths, configurations, and governance implications.

Option 1: Smart Tags

Smart Tags use Adobe Sensei (AEM's AI/ML framework) to automatically analyze image content and apply tags based on visual recognition. Smart Tags work particularly well for image-heavy asset libraries where visual attributes—color, composition, subject matter—need to be surfaced in search.

Key characteristics:

TIP
Training Smart Tags is not a one-time event - it is an ongoing governance responsibility. DAM teams must regularly supply high-quality training sets, review Smart Tag output, and feed rejected tags back into the system as negative training examples. Without this feedback loop, the model's confidence scoring will drift over time and produce increasingly irrelevant tags.

Option 2: Enhanced tagging (AI-generated metadata)

AEM Assets as a Cloud Service also offers Enhanced Tagging via the Cloud Services AI-generated metadata capability. While Smart Tags focus on visual analysis, Enhanced Tagging takes a broader, contextual approach - analyzing the asset's associated metadata, descriptions, naming conventions, and information hierarchy to generate richer, more semantically relevant tags and metadata fields.

Key characteristics:

How taxonomy structure becomes a dependency for Enhanced Tagging

Enhanced Tagging doesn't operate in a vacuum. When the AI analyzes an asset, it cross-references available metadata against your taxonomy to determine which controlled terms best apply. A folder containing assets with consistent naming conventions, well-formed descriptions, and correctly applied parent taxonomy terms gives the AI clear, unambiguous signals. A folder with inconsistent naming, missing descriptions, and orphaned or duplicate taxonomy nodes gives the AI contradictory signals—and the resulting tags will reflect that chaos.

Your taxonomy is the training environment for Enhanced Tagging.

Choosing the right approach

Most mature organizations will use both capabilities in concert. Smart Tags handle the visual layer; Enhanced Tagging handles the semantic and contextual layer. The governance model you design must account for both.

Capability
Best for
Primary signal
Governance lever
Smart Tags
Image-heavy libraries, visual attribute search
Pixel/visual recognition
Training sets, confidence thresholds
Enhanced Tagging
Contextual discovery, semantic search, multi-asset types
Metadata, context, hierarchy
Folder-level settings, approval workflows

The AI solution: context-driven automation

We cannot change how tags work in AEM Assets or how they facilitate the page experience, but we can use AI as an accelerator to scale and strengthen the tagging process.

Rather than relying on a simple keyword search, AI can detect the true context of an asset by analyzing the page's URL, description, keywords, and overall information hierarchy. AI then maps this output directly to the organization's controlled taxonomy.

●      Confidence Scoring: When a common keyword exists in multiple places across the taxonomy, the AI analyzes the page or asset’s context to assign a confidence score to each potential option, ensuring the most relevant iteration of the tag is applied.

●      Smart Selection: Instead of blindly applying a generic "Horizon" tag, the AI is smart enough to see the context and apply "Horizon Summit events" over "Horizon Summit campaign.”

This is not automation replacing judgment. It is automation operationalizing judgment at a scale no human team can maintain alone.

Designing a hybrid governance model

The future of asset management isn't a competition between human and machine. It's a partnership. AI takes the first stab at contextualizing and tagging an asset based on the foundation the DAM team provides. Again, AI tagging is a strong partner, but a human-in-the-loop is still required to ensure the AI is trained and continues to learn correctly.

To safely deploy this at scale, organizations must design a hybrid governance model:

Ambiguous, manual tagging creates chaotic metadata systems that have the potential to hurt and weaken discovery and personalization. By implementing a hybrid governance model, organizations can leverage AI to solve complex contextual challenges.

How it works in practice

To make this concrete, here is how the hybrid governance model plays out for a new asset batch entering the DAM.

Step 1 - Asset ingestion: A batch of event photography from the Horizon Summit is uploaded to the designated folder. The folder is pre-configured with Enhanced Tagging enabled and a confidence threshold of 80%.

Step 2 - AI tagging runs: AEM Assets analyzes each image. For a photo of the keynote speaker on stage, the AI evaluates:

The AI maps these signals to the controlled taxonomy and suggests:

Step 3 - Librarian review: The DAM Librarian opens the review queue. They see the keynote tag suggestion at 72% confidence. They confirm it is correct, approve the tag, and it is applied. The approved match is logged as a positive training example.

Step 4 - Bulk metadata augmentation: The librarian notices that the entire batch is missing the Distribution > Internal Use Only tag. They select all 47 assets in the batch and apply the tag in bulk - a single action rather than 47 individual edits.

Step 5 - Feedback loop: Two suggestions in the batch were rejected: the AI incorrectly suggested Campaign > Product Launch for a general summit photo. The librarian rejects both and logs them as negative training examples. Those rejection signals are fed back to the model to reduce similar misclassifications in future batches.

The foundation still comes first

All of this requires a bottom-up approach. AI is only as smart as the data it learns from. The cleaner and more accurate your foundational metadata, controlled vocabularies, and keyword structure, the more effective your AI tagging will become.

In our personal experience, we are treating it as a digital clean slate. It is tempting to simply lift and shift existing taxonomies, assets, and metadata and flip the switch on, but we recognized early on that doing so would just automate legacy inconsistencies. We are deliberately holding off on turning on AI tagging until our house is in order, ensuring our move to the cloud is backed by accurate data, terms, and vocabulary.

Before deploying AI-assisted tagging, ask:

If the answer to any of these is no, that is where to start. AI applied to a weak foundation will scale your inconsistencies faster than any team of humans ever could.

Try it

Follow these steps in building a human and AI hybrid governance model:

Key takeways

Resources

style
article-content-section