The contract governing data moving between organisations. It sets out what may be done with it, for how long, by whom, with what security, and what happens at the end. For AI work the clause everyone forgets is whether the recipient may use it to train a model, and whether the resulting model is then constrained.
Technical view
Should specify purpose limitation, permitted processing including whether model training is in scope, sub-processor rights, security requirements, audit rights, breach notification, retention and deletion including derived artefacts such as embeddings and weights, and the position on ownership of outputs. International transfers require an appropriate mechanism. Model provenance obligations increasingly appear, requiring the recipient to record which datasets contributed to which model version.
Related
Part of AI Terminology Index, section: Data.