Company data quality determines the outcome of an AI deployment more than the choice of model. Companies only discover this during the pilot, when the solution answers incorrectly and the cause is not the technology but the underlying data. The most common problems are fourfold: the same customer or product recorded multiple times in different forms, missing values in fields that decisions are meant to be based on, contradictory data between systems, and documents in a form that cannot be processed, meaning scans without recognised text and tables inserted as images. Before deployment, therefore, carry out a brief inventory of the data the solution is meant to use, and clean up at least the data the result depends on. Include data preparation in the project budget – it tends to be its largest item.
See also: Data audit, Data normalisation, RAG over a company's own content.
