Data onboarding starts when a new source or a new customer’s data arrives and ends when that data can be used for the job it was brought in for. The steps are usually the same: connect to the source, map its fields to the names and types the receiving system expects, check for gaps, duplicates, and inconsistent formats, and confirm with the owner what each field means and which records are exceptions. In advertising, the same term has a narrower meaning, matching offline customer records to online identifiers; this entry covers the general case.
For AI, loading the data is the midpoint. Onboarding ends when the data has been checked against the use case it will serve: the fields the model needs are present, their meaning is confirmed, and a result on this data can be compared with a known baseline. Much of this work is done by hand today, often by a forward deployed engineer, and it repeats for every customer. Recording the mappings, preparation rules, and acceptance criteria as reusable assets turns onboarding from a project into a repeatable step toward AI-ready data.