Many enterprises encounter a reality when implementing ERP/digitalization: the more data they have, the more unstructured content there is (documents, images, voice, work orders), but only a limited amount can actually be used by the business. Recently, Google Cloud introduced a set of AI functions that integrate Gemini/Vertex AI capabilities more "natively" into BigQuery, allowing data teams to use SQL for tasks such as summarization, extraction, translation, structured output, vectorization, and similarity search. This approach is highly insightful for enterprise data platforms: turning AI into a standard function of the data layer, rather than scattering it across various scripts and applications.
What value does this have for ERP/digitalization?
- Lower the barrier: Business analysts can directly call AI within their familiar SQL environment, instead of learning a new set of tools.
- Making unstructured data "computable": Transforming text/image understanding and structured extraction into table fields and metrics that can be incorporated into processes such as reporting, risk control, and supply chain forecasting.
- Easier to govern: When AI calls become part of the data platform, permissions, auditing, and cost accounting are more easily integrated into the corporate governance system.
3 actionable use cases (we suggest you start here)
- Work Order/Contract Summary + Structured Fields: Convert long text into fixed fields (customer, amount, risk points, delivery date) and directly populate the table.
- Similar Issue Retrieval (Semantic Search): Use embedding + similarity / vector search to quickly find "similar tickets/similar faults/similar clauses".
- Multi-task generation in one go: Perform entity extraction, sentiment/risk assessment, translation, and summarization simultaneously with a single SQL call, outputting multiple columns of results that directly enter the data pipeline.
Implementation Notes (Avoid Pitfalls)
- Permissions and Compliance: Design "who can call, which models can be called, and which data can be seen by the model" as first-class citizens.
- Cost and Caching: Prioritize batch processing/incremental updates for high-value scenarios; do not make real-time calls to the large model for all queries.
- Quality Evaluation: Establish a sampling acceptance mechanism for summarization/extraction; key fields must be able to trace back to original text evidence.
References
The views in this article are based on secondary collation and expansion of public materials. It is recommended to read the original text for details:
Google Cloud Blog: New BigQuery gen AI functions for better data analysis
