The documented record of where a piece of data or a generated asset came from, who made it, what was done to it, and what it is licensed for.
Also known as data lineage, content provenance, chain of custody
Data provenance is the traceable history of a dataset or an asset: its origin, its chain of custody, the transformations applied along the way, and the terms it can be used under. On the training side, that means knowing which corpora a model learned from and whether those corpora were licensed. On the output side, it means being able to say which model produced an image, from what prompt, with what reference material, and on what date.
Provenance is documentation, not detection. A signed manifest travels with a file and asserts a verifiable claim about its history. A classifier guesses after the fact and is far less reliable. The practical difference matters more than it sounds: provenance is established at the moment of creation, which means the time to set it up is before the asset exists, not when a client’s counsel asks about it.
Provenance moved from a governance abstraction to a contract question faster than most studios adjusted.
Client contracts increasingly ask for it. Master services agreements now routinely carry AI usage warranties: a representation that delivered work does not infringe, and a disclosure of what was machine-generated. You cannot warrant what nobody recorded.
Disclosure rules now have teeth. The EU AI Act’s transparency obligations became fully enforceable in August 2026, and several US states have added their own labeling requirements. If a campaign runs in those markets, somebody on the team has to know which assets are synthetic and be able to show it.
It protects the studio, not only the client. When a dispute lands over whether an image was generated from someone else’s photography, the studio with a logged prompt, a licensed reference, and a model version settles it in an afternoon. The studio without one spends a month reconstructing a year-old decision from Slack.
A brand team asks for source files on a key visual eleven months after the campaign ran, because a stock library has sent a letter. The retoucher who built it has left. With provenance logged, the answer takes twenty minutes: the base plate came from the client’s own shoot, the sky replacement was generated by a named model on a named date, and the reference board held only licensed material. Without it, the studio is guessing in public. The agencies that get this right tend to do it the boring way, with a required field in the asset manager and a rule that no generated asset ships unless its prompt and model version ship with it.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.