Ten thousand customer reviews. Hundreds of support tickets. Product descriptions written in different languages. A growing collection of documents that contains valuable business information, but doesn’t fit neatly into a dashboard.
Most organizations already have this data. The harder question is how to make sense of it.
Traditional SQL is excellent at filtering transactions, calculating revenue, and identifying trends in structured datasets. But what happens when the answer is hidden inside a sentence, an image, or the meaning of a customer complaint?
That’s where BigQuery AI functions change the conversation. Google Cloud is making it possible to bring generative AI and semantic analysis directly into BigQuery SQL workflows. Instead of exporting data to a separate application for every AI task, data teams can use SQL to summarize content, extract information, generate embeddings, and identify semantic relationships.
Recent updates to BigQuery generative AI functions make these capabilities easier to use, more flexible, and more manageable at scale.
Let’s look at what has changed, how the functions work, and where they can deliver practical value.
The SQL Query That Can Read Between the Lines
Imagine an e-commerce company with thousands of product reviews.
A conventional query can calculate the average rating, count negative reviews, or identify the most frequently purchased products.
But suppose the marketing team wants to know:
What are customers actually complaining about?
Some customers mention delivery delays. Others describe confusing product instructions or packaging problems. Many reviews contain several different issues in the same paragraph.
Searching for specific keywords will miss complaints expressed in unexpected ways.
With AI-powered SQL, analysts can ask a language model to interpret the review text and return structured information that can be analyzed alongside existing business data.
The important development is that this can happen within a SQL query.
AI.GENERATE: Turning Unstructured Content into Useful Data
Among the latest BigQuery AI functions, AI.GENERATE is one of the most versatile.
It enables data teams to perform language-model inference directly in SQL, including summarization, translation, information extraction, and classification.
In January 2026, Google Cloud announced that AI.GENERATE and AI.GENERATE_TABLE had reached general availability.
The release also expanded support for structured and unstructured inputs, including text and supported image, audio, video, and document data.
One Query, Multiple Insights
Consider a table containing customer feedback.
An analyst wants to identify the sentiment, main issue, and a short summary for each review.
Rather than building three separate processing workflows, AI.GENERATE can return multiple structured fields from one inference call.
A simplified query could look like this:
SELECT
review_id,
AI.GENERATE(
review_text,
output_schema =>
‘sentiment STRING,
main_issue STRING,
short_summary STRING’
).*
FROM `my_project.retail.customer_reviews`
LIMIT 100;
Illustrative example: Replace the table and column names with your own. The query assumes the necessary model access and permissions are configured.
The output_schema parameter tells BigQuery which fields to return.
Instead of receiving a free-form paragraph that requires additional processing, the analyst gets values that can be used in subsequent SQL operations.
This makes the function particularly useful for:
- Customer feedback analysis and issue classification.
- Document summarization and information extraction.
- Multilingual content processing.
- Support ticket categorization.
- Product catalog enrichment.
The output still needs validation. A language model can misinterpret ambiguous feedback or generate inaccurate classifications, particularly when the source material is incomplete.
Nevertheless, AI.GENERATE creates a practical connection between unstructured content and familiar analytical workflows.
AI.EMBED: Making Meaning Searchable
Not every business question can be answered by generating text.
Sometimes the goal is to find information that means something similar, even when the wording is completely different.
Consider these two customer requests:
“My package still hasn’t arrived.”
“Delivery is taking much longer than expected.”
A traditional keyword search may treat them differently. A semantic system can recognize that both concern delivery delays.
This is where AI.EMBED becomes useful.
The function converts supported content into embeddings: numerical vectors that represent semantic information.
Content with similar meaning can have embeddings that are close to one another in vector space.
For example:
SELECT
review_id,
AI.EMBED(
review_text,
endpoint => ‘text-embedding-005’
).result AS review_embedding
FROM `my_project.retail.customer_reviews`
LIMIT 100;
The result is a vector representation of each review.
These embeddings can support semantic search, recommendations, clustering, and other machine learning workflows.
Why Embeddings Matter for Enterprise Analytics
Organizations often manage information that is difficult to categorize using fixed rules.
A support request may describe a technical problem without using the official product terminology. A product description may be relevant to a search even when it doesn’t contain the exact search phrase.
Embeddings help bridge that gap.
For example, a retailer could use AI.EMBED to develop product discovery experiences based on descriptions and meaning rather than exact keyword matches.
A support team could group similar requests to identify recurring problems.
A knowledge management team could use embeddings to help retrieve relevant documentation for an AI application.
For large-scale search systems, embeddings can also be stored and used with BigQuery’s vector search capabilities.
The key is choosing an approach appropriate to the size and performance requirements of the workload.
AI.SIMILARITY: Finding Connections Beyond Keywords
While AI.EMBED creates vector representations, AI.SIMILARITY offers a more direct way to compare the semantic meaning of two inputs.
The function generates embeddings and calculates their cosine similarity.
This allows analysts to compare text without first building a separate embedding pipeline.
Imagine a customer service team searching for tickets related to delivery delays.
Instead of searching only for the word “delivery,” an analyst could compare a description of the issue against ticket text.
SELECT
ticket_id,
ticket_description,
AI.SIMILARITY(
‘Customer experiencing a delayed shipment’,
ticket_description,
endpoint => ‘text-embedding-005’
) AS similarity_score
FROM `my_project.support.tickets`
ORDER BY similarity_score DESC
LIMIT 10;
The query ranks tickets by semantic similarity to the requested issue.
This can reveal relevant records even when they use different vocabulary.
When Should You Use AI.SIMILARITY Instead of Vector Search?
The distinction matters because both approaches can support semantic retrieval.
| AI.SIMILARITY | VECTOR_SEARCH |
| Compares content directly | Searches collections of embeddings |
| Generates embeddings during evaluation | Typically uses precomputed embeddings |
| Useful for interactive analysis and smaller comparisons | Better suited to large-scale nearest-neighbor retrieval |
| Does not use vector indexes | Can benefit from vector indexing |
For an analyst exploring a relatively small dataset, AI.SIMILARITY can be a convenient starting point.
For a production search application operating across millions of records, precomputed embeddings and VECTOR_SEARCH may offer a more efficient architecture.
Choosing the right method can have a significant impact on both performance and cost.
What’s New Beyond the Three Core Functions?
The latest developments extend beyond individual SQL functions.
Google Cloud is gradually making generative AI a more integrated part of analytical workflows.
1. Simplified Authentication with End User Credentials
Previously, calling Vertex AI models from BigQuery often required configuring a connection and managing service account permissions.
Google Cloud introduced support for End User Credentials (EUC), allowing eligible interactive queries to use the user’s IAM identity.
This can simplify initial setup for supported workflows, provided the required Vertex AI permissions are available.
For data teams experimenting with AI SQL functions, fewer configuration steps can make prototyping more accessible.
However, organizations still need to evaluate permissions, production execution identities, and governance requirements.
2. Gemini Model Support in BigQuery
The integration of Gemini in BigQuery allows analysts to apply supported Gemini models to data processing tasks through SQL.
Google Cloud’s January 2026 update introduced Gemini 3.0 Pro and Flash support for generative AI inference.
Model selection matters because different workloads have different requirements.
A straightforward text classification task may prioritize throughput and cost efficiency. A more complex reasoning task may require a different model.
The most effective approach is to evaluate output quality, latency, availability, and cost against the actual use case.
3. More Specialized AI Functions
The broader BigQuery generative AI functions portfolio also includes managed functions designed for common analytical tasks.
Examples include:
- AI.CLASSIFY for categorizing content.
- AI.SCORE for evaluating or ranking inputs.
- AI.IF for filtering data using natural-language conditions.
- AI.AGG for AI-assisted aggregation and summarization.
These functions can reduce the need to write detailed prompts for every common task.
For teams already comfortable with SQL, this creates an opportunity to incorporate AI-assisted analysis into existing workflows without developing an entirely separate application.
Feature availability, model support, and optimization options should be checked for the intended deployment.
The October 2026 Update: Better Control Over AI Costs
There is another development worth paying attention to, especially for enterprise data teams.
On October 14, 2026, Google Cloud announced the general availability of token-based cost management for BigQuery generative AI workloads.
This addresses an important operational challenge.
When an AI function runs across thousands or millions of rows, each model invocation can consume tokens.
A query that looks simple in SQL may therefore generate substantial AI inference costs.
The new cost management capabilities allow administrators to configure daily token quotas for supported generative AI functions.
These quotas provide guardrails that help teams control AI consumption.
This is particularly relevant for organizations expanding AI-powered SQL from experimentation into recurring production workloads.
A Practical Approach to Cost Management
Before running AI inference across a large dataset, data teams should consider three questions.
Does every row need AI processing?
Often, conventional SQL can narrow the dataset before invoking a model.
For example, filtering reviews by date, category, or rating may substantially reduce the number of inference calls.
Can the results be reused?
If the underlying content rarely changes, repeatedly generating the same classifications or embeddings may be unnecessary.
Storing validated outputs can reduce repeated processing.
Is the selected model appropriate for the task?
More capable models are not automatically the most economical choice for every workload.
Testing different models against quality and cost requirements can help identify an appropriate balance.
Token quotas add an important layer of control, but they do not replace query optimization, workload monitoring, or financial planning.
Where BigQuery AI Functions Can Deliver Business Value
The most useful applications are often found where organizations already have large amounts of unstructured data.
Retail: Understanding Customer Feedback
Retailers can use AI.GENERATE to extract recurring complaints and product improvement opportunities from reviews.
By combining those results with structured sales data, teams can investigate whether specific customer concerns correlate with returns, ratings, or purchasing patterns.
Financial Services: Organizing Complex Documents
Financial institutions process substantial volumes of textual information, including reports, customer correspondence, and operational documentation.
BigQuery generative AI functions can assist with extracting relevant fields, summarizing documents, and organizing information for downstream analysis.
Sensitive workflows still require appropriate data protection, quality checks, and human oversight.
Customer Support: Finding Similar Cases
Support organizations can use AI.SIMILARITY to identify tickets that describe related problems.
This can support issue discovery, knowledge retrieval, and the identification of recurring service problems.
Media and Content Platforms: Improving Discovery
Media organizations can use AI.EMBED to represent content semantically.
These embeddings can support recommendation systems, content grouping, and search experiences that go beyond exact keyword matching.
Across these examples, the opportunity is not simply to add AI to an existing query.
It is to make previously difficult-to-analyze information more accessible to the business.
The Bigger Picture: SQL Is Becoming an Interface to AI
The development of BigQuery AI functions reflects a broader change in how organizations work with data.
Historically, structured analytics and AI processing often followed separate workflows.
SQL handled reporting and data transformation. Specialized applications handled natural language processing, embeddings, and model inference.
With AI-powered SQL, those capabilities are becoming more closely connected.
An analyst can combine traditional calculations with AI-generated insights in a single analytical workflow.
A data engineer can enrich existing datasets with structured information extracted from text.
A development team can use embeddings to support semantic retrieval and AI applications.
This doesn’t mean every analytical task should involve a language model.
Traditional SQL remains the right choice for deterministic calculations, aggregations, and many reporting requirements.
The value comes from knowing when AI adds something that conventional data processing cannot easily provide.
Turn AI-Powered SQL into Measurable Data Value with Oredata
The latest BigQuery AI functions make it easier to explore generative AI directly within enterprise data workflows.
But moving from an interesting SQL experiment to a reliable production process requires more than knowing the syntax.
Organizations need to consider data quality, model selection, access controls, performance, inference costs, and how AI-generated results will be validated and consumed.
As a Google Cloud Premier Managed Service Provider, Oredata helps enterprises modernize their data platforms, optimize BigQuery workloads, and develop scalable data and AI solutions.
Our focus is on connecting technology decisions with practical business outcomes, whether that means improving analytical performance, making unstructured data more useful, or designing AI-enabled data workflows.
Your Data Already Contains the Questions. Let’s Make the Answers Easier to Find.
Choose one dataset your team struggles to analyze today: customer feedback, support tickets, documents, or product content.
Oredata can help you evaluate which BigQuery AI functions are appropriate, what the implementation requires, and how to measure the value of the results.
Explore what’s possible with your existing data. Talk to Oredata’s Data & AI team about building an AI-powered analytics use case around your business priorities
