Short answer
Auto-tagging is the automatic assignment of descriptive metadata to assets by AI image, video or text analysis. Auto-tagging works best when constrained to the controlled vocabulary. Without that constraint, generic labels such as 'outdoor' or 'person' add search noise, while brand, product and rights fields stay empty and still need human input.
How auto-tagging works in a DAM
Auto-tagging works by running each new asset through one or more AI models on ingest and writing the results to metadata fields. Each suggested tag usually carries a confidence score, and the DAM applies rules to that score.
A typical auto-tagging setup has five parts:
- Analysis: image recognition, optical character recognition, speech-to-text for video, or text extraction for documents.
- Mapping: model labels map to values in your controlled vocabulary.
- Threshold: tags at or over a set confidence apply automatically, and tags under it go to a review queue.
- Labelling: the DAM marks which tags came from AI, so users and audits can tell them apart.
- Custom training: some DAMs train a model on your own products, colours or logos.
Auto-tagging consumes AI credits or compute per asset. DAMSelection assessment: storage, egress and AI credit consumption drive five-year DAM cost more than licence fees, so price every tagging run, including re-tagging after a model change.
Example of auto-tagging
Worked example: a fashion retailer ingests 10,000 product images per season. An AI model tags each image with colour and material values from the schema's controlled lists.
Tags over the confidence threshold apply directly. Tags under it land in a review queue, so the librarian checks only the uncertain images. Product SKU and rights expiry still arrive from the PIM and the rights workflow, because no image model can read a licence.
Auto-tagging in a DAM selection
Auto-tagging is proven at Checkpoint 5: Test, on your own assets. DAMSelection assessment: generic vendor demos carry no selection value, and stock photos make every tagging model look accurate.
Ask each vendor to run these tests:
- Tag 200 of your own images and compare the result with tags from your librarian.
- Restrict output to your controlled vocabulary for at least one field.
- Show the confidence threshold setting and the review queue.
- Show how AI tags are labelled in the interface and the API.
- State where the model processes your files and how credits are counted per asset.
The AI in DAM guide explains model types. The TCO calculator adds AI credit consumption to the five-year cost.
Common auto-tagging mistakes
The most common auto-tagging mistake is judging accuracy on the vendor's sample images. Other recurring errors:
- Letting AI write rights, product or legal fields that need a source record.
- Leaving AI tags unlabelled, so nobody can audit or remove them later.
- Ignoring the cost of re-tagging the full library when the vendor changes models.
- Accepting every generic label, which floods search with low-value tags.
Related terms: Semantic search, Controlled vocabulary.
Frequently asked
How accurate is DAM auto-tagging?
DAM auto-tagging accuracy depends on your images, your vocabulary and the model, so no general figure applies. Measure it in a [proof of concept](/dam/select/proof-of-concept/): tag a sample of your own assets and compare the result with a librarian's tags. Score the share of correct tags per field.
Does auto-tagging replace a DAM librarian?
Auto-tagging reduces manual tagging of descriptive fields such as colour, object and scene. A librarian still owns the vocabulary, reviews low-confidence tags and audits AI output. Product, rights and campaign fields still come from source systems and people.