Choose generative AI in adtech only after a stack check
The decision: first the decisioning layer, then the creative layer
Choose generative AI only for a specific advertising workflow where the existing decisioning layer can already be assessed. The supplied NIST passage requires the intended use and business context to be defined or, for existing systems, reassessed. It also identifies classification, generative models and recommenders as distinct tasks and methods. This is not evidence that a particular adtech stack performs better; it is a useful design boundary. Before running a creative pilot, therefore, establish which task the model supports, which product and contextual information precedes it, who uses the output and where human oversight must remain.
Privacy and risk belong in the same decision
Do not treat privacy as a final check on generated assets. In the NIST passage, system requirements, with privacy as an example, sit alongside documented risk tolerances and socio-technical design decisions. For an adtech team, this means in practice: determine for each workflow which data and signals the generative layer may receive, who oversees it, and at which deviation the workflow stops or falls back to the existing route. The passage does not prescribe a specific privacy design or assess an advertising platform; it only supports the need to make requirements and risk trade-offs explicit in advance.
Make the alignment verifiable in a single register
Do not use a broad AI roadmap as the first step; instead, assess one workflow from input to decision. Complete a stack check covering: workflow and task; owner; permitted input and privacy requirement; the existing ranking, matching or measurement step the output connects to; risk limit; and stop or fallback decision. For providers of general-purpose AI models, the supplied Commission passage mentions, among other things, technical documentation and information for downstream AI system providers about capabilities and limitations. According to that passage, this applies to providers of such models under the AI Act, not automatically to every adtech team. Still, it is a relevant reason to explicitly request this information from a supplier before the creative layer enters an operational workflow. Practical tool — stack check for one advertising workflow: record the supported task, the owner, the permitted input, the privacy requirement, the linked decision step, the checkpoint and the fallback decision. Limitation: the supplied passages do not substantiate claims about conversion, auction latency, embeddings, real-time bidding or the legal obligations of individual advertisers. Have the applicable privacy, contractual and AI Act obligations for your own role and setup assessed separately.
