Use case assessment
Is AI image tagging worth building?
Genuinely AI, thoroughly commoditised, and there's one variant that changes everything.
Real AI, mature and cheap. Unless you are tagging people, in which case you have left ordinary image tagging and entered biometrics.
What it usually means
Automatically labelling images with what they contain, for search, organisation or accessibility.
Is it genuinely AI?
Yes, and it is among the most commoditised capabilities available. Object and scene recognition is accurate, fast and priced low enough that building your own requires a specific reason.
The variant that changes everything
Tagging objects is one thing. Tagging people is another.
Facial recognition and biometric categorisation are treated entirely differently under the EU AI Act. Biometric identification systems sit in Annex III high-risk. Some biometric categorisation — inferring characteristics such as race, political opinion or sexual orientation — is prohibited outright. And untargeted scraping of facial images to build recognition databases is also prohibited.
Many general-purpose vision services include face detection and grouping. Turning that on in a digital asset manager, for convenience, moves you across a line that "image tagging" does not suggest exists.
What the simpler version looks like
Structured upload metadata. If your team can tag at upload with a controlled vocabulary, that is more accurate than inference and searchable by terms your organisation actually uses.
Automatic tagging earns its place on a backlog of untagged assets, or where upload volume makes manual tagging impractical.
What it costs to run
Per-image pricing, low and predictable. The hidden cost is vocabulary alignment: generic tags rarely match how your organisation describes things, so you end up mapping model output to your own taxonomy, and that mapping needs maintenance.
Where it sits under the EU AI Act
Object and scene tagging: not Annex III. Face detection, recognition or grouping: a different regime entirely, and one to decide deliberately rather than enable.
Images of people are personal data under GDPR whether or not you tag them, so retention and lawful basis apply regardless.
When it is worth it
Buy. Worth adopting at volume, on a large untagged archive, or where accessibility alt text is needed at a scale humans cannot cover.
With face features explicitly reviewed and, in most cases, off.
Common questions
Should we build our own image tagging?
No. Object and scene recognition is thoroughly commoditised, accurate and cheap. Building it requires a specific reason that most organisations do not have.
Is image tagging regulated under the EU AI Act?
Object and scene tagging is not Annex III high-risk. Face detection, recognition or grouping is a different regime: biometric identification is high-risk, some biometric categorisation is prohibited, and untargeted facial image scraping is prohibited.
What is the alternative to automatic tagging?
Structured metadata at upload with a controlled vocabulary. More accurate than inference and searchable by the terms your organisation actually uses. Automatic tagging earns its place on backlogs and at high upload volume.
Scoring 40 AI use cases in a regulated enterprise — how one enabled feature moved a low-risk tool into a different regime.