Reference Glossary
AI auto-tagging
Automatic assignment of descriptive keyword tags to assets by a machine learning model trained on labelled images, without a person manually typing each tag.
Why it matters in a DAM
For a legacy library of tens of thousands of untagged assets, auto-tagging is often the only realistic way to get any baseline searchability — manually tagging that backlog just isn't going to happen on any reasonable timeline. Cloudinary, for instance, lets a DAM call Amazon Rekognition or Google Vision at upload time to assign scene-category tags automatically, with a confidence threshold controlling which suggested tags get kept.
A worked example
Common mistake
Teams pipe raw AI-generated tags straight into the search index without mapping them to the DAM's own controlled vocabulary, so the library ends up searchable by generic model output like 'outdoor', 'person', or 'text' that doesn't match how anyone on the brand team actually searches.
AI auto-tagging hands the tagging job to a machine learning model instead of a person: upload an image, and the model returns a list of labels it recognises in it, with a confidence score attached to each. In a DAM, those labels typically get written into the same tag or keyword field a human would have filled in by hand.
The case for it is mostly about scale. A library with a ten-year backlog of untagged photography can’t realistically be caught up by manual review, so auto-tagging becomes the difference between that backlog staying invisible to search forever and it being at least partially findable. Cloudinary‘s integration is a concrete example of how this is wired in practice: setting a categorisation parameter on upload routes the image through Amazon Rekognition or Google Vision, and the returned scene categories get attached as resource tags automatically.
What auto-tagging doesn’t do on its own is speak the library’s language. A generic vision model returns generic categories — ‘person’, ‘outdoor’, ‘vehicle’ — because that’s what it was trained to recognise, not because it knows the brand’s product names or campaign terminology. Without a mapping step from model output to the DAM’s actual controlled vocabulary, auto-tagged assets end up searchable in a vocabulary nobody on the team uses.
Frequently asked
What does AI auto-tagging do, and how is Cloudinary's implementation a concrete example?
A machine learning model automatically assigns descriptive keyword tags to an asset without a person typing them; Cloudinary can route an uploaded image through Amazon Rekognition or Google Vision at upload time to assign scene-category tags, with a confidence threshold controlling which suggested tags get kept.
Why is auto-tagging often the only realistic option for a large legacy library?
Manually tagging a backlog of tens of thousands of untagged assets just isn't going to happen on any reasonable timeline, so auto-tagging becomes the difference between that backlog staying invisible to search forever and being at least partially findable.
What's the risk of piping raw AI-generated tags straight into the search index?
The library ends up searchable by generic model output like 'outdoor', 'person', or 'text' that doesn't match how anyone on the brand team actually searches, because the model wasn't trained on the brand's own vocabulary.
Does AI auto-tagging know a company's product names or campaign terminology?
No — a generic vision model returns generic categories because that's what it was trained to recognise, not because it knows brand-specific names; a mapping step from model output to the DAM's own controlled vocabulary is needed to close that gap.
What role does a confidence threshold play in AI auto-tagging?
Each tag the model returns carries a confidence score, and the threshold is the cutoff deciding which scores are trusted enough to write to the asset record automatically — as in Cloudinary's Amazon Rekognition/Google Vision integration at upload time. Tags below that threshold aren't necessarily discarded; they can instead be routed to manual review rather than auto-accepted, since a low-confidence tag is more likely to be inaccurate.
What's an example of AI auto-tagging output being mapped to a brand vocabulary?
Model output such as 'outdoor, person, footwear, blue' — generic categories a vision model recognizes — gets mapped to a brand category path like 'Lifestyle photography > Product-in-use > Footwear' in the DAM's controlled vocabulary. That mapping step is what closes the gap between what the model actually detects and the terms the brand team already searches by, since raw model output alone doesn't carry brand-specific names.
Sources
- Cloudinary can route uploaded images through Amazon Rekognition to automatically classify scene categories and assign them as resource tags, controlled by a confidence threshold. checked 2026-08-07 — Cloudinary Documentation — Amazon Rekognition Auto Tagging Add-on