This article is published as part of Google Cloud Next ‘26 and is co-authored by Manan Goel, Group Product Manager, Databases and Analytics
AI is changing how we use data. Instead of just storing information for us to look at later, data platforms are becoming “active” partners that can take real action. Modern data platforms are evolving from “system of intelligence” to “system of action” in three main ways:
-
From operating at human scale to agent scale: We are moving from humans asking a few questions a day to AI agents asking thousands of questions every minute. This means our systems must be much faster, scalable, and more affordable to keep up.
-
From reactive intelligence to proactive action: Instead of waiting for us to check a dashboard, AI can now spot trends and act on them in real-time.
-
From data to semantic knowledge: Traditional data platforms use rigid schemas and stale metadata. AI agents use “semantic search” to understand how different pieces of information relate to one another, much like a human would.
From Prototypes to Production
Many companies are excited to start using AI agents this year, but moving from a demo to a real-world system is a big step. A reliable agent shouldn’t just wait for a command like a chatbot; it should be a proactive companion. It needs to break down big goals into small steps, remember the context of its work, and—most importantly—follow security, governance, and compliance rules while constantly learning from feedback.
Making it Work with BigQuery
You can start using agents in BigQuery with two main ways. You can use one of the pre-built agents like the Data Engineering Agent to help automate data pipeline development or Conversational Analytics to ask questions in plain language. Alternatively you can use custom tools such as APIs, Agent Development Kit (ADK), and MCP to build your own custom agents. For the purposes of this blog we’ll focus on the Data Engineering and the Conversational Analytics agents within BigQuery.
Imagine you’re an e-commerce data engineer during the peak holiday season. You need to build a pipeline where a Data Agent automatically ingests real-time inventory updates from a Cloud SQL PostgreSQL database, parses nested JSON vendor logs in BigQuery, and dynamically masks customer shipping addresses (PII) before the analytics team builds their dashboard.
Ingest, prepare and transform this data with the Data Engineering Agent
Step 1: Integrate Public Data
Subscribe to a relevant public dataset from the BigQuery Analytics Hub (such as Google Trends or a public retail dataset).
Once it’s linked to your project, invoke the agent by selecting ‘Ask Agent’ in BigQuery Pipelines. Instead of manually writing complex joins, give the agent a specific, actionable prompt:
“Create a plan for a data pipeline that joins our internal holiday_sales table with the Analytics Hub public_retail_trends dataset on the date and zip_code columns, and outputs the enriched data into a new table called holiday_inventory_forecast. Allow me to review the plan first.”
Step 2. Clean and Organize: Once your raw data is available, you can get the data ready for analysis. This means handling the messy data and the security requirements. Treat the agent like a junior developer and ask it to review the spec first:
“Parse the nested JSON vendor_logs column in the holiday_inventory_forecast dataset to extract item_id and supplier_lead_time. Then, apply dynamic data masking to the shipping_address column using our standard pii_masking_policy. Outline the exact steps you plan to take before executing.”
Empowering the Business with Conversational Analytics
Now that your data is clean, enriched, and masked, you can set up Conversational Analytics so business teams can get insights without writing SQL.
Step 1. Contextualize the Agent: Open BigQuery, navigate to Agents, and create a new agent pointed directly at your new holiday_inventory_forecast tables. Give the agent vital business context in the instructions. For example:
“Interpret ‘low stock’ as any stock_level below 50 units. If asked about regions, map the standard US state abbreviations to their respective geographic regions.”
Click Publish and share it with your analyst group.
Step 2. Have Analysts Access the Agent: Your analysts can now go to the Agent Catalog under “Shared by others in your organization” and start querying immediately. They don’t need to know the underlying table structures; they can just ask:
“Which products in the northeast region are currently categorized as ‘low stock’ based on the new forecast?”
Try it yourself
Don’t want to spin this up in your own environment? Try these agents in a Google Cloud Skills Boost interactive lab: Build and Evaluate BigQuery Agents using Agent Development Kit (ADK) and GenAI Eval Service. For instructions on the Data Engineering Agent, follow this blog post with step by step instructions. For instructions on Conversational Analytics, follow this blog post.
Developing a Clear Spec First
Many new AI coders fall into a common trap. They give the AI vague instructions and let it start writing code right away. This approach leads to errors, bugs, and wasted time.
Before generating any code, consider creating a clear specification. This plan tells the AI exactly what the software should do, how it should work, and what limits it has. Iterate with your agent on the spec, not on the code.
When an AI agent receives vague directions, it makes guesses. These guesses often do not match your goals. If the agent writes a lot of code based on incorrect guesses, fixing it takes more work than writing it correctly the first time. The agent might also choose the wrong tools or build a system that is hard to maintain.
Iterating on the Spec Treat the spec as a working document. You should work with the AI to improve the spec before you ask it to write any code. Follow these steps to build a solid plan:
-
Draft the initial idea: Write down what you want to build in plain text. Include the main goal and the required features.
-
Ask the AI for feedback: Have the AI review your idea. Ask it to point out missing details, logic gaps, or potential errors.
-
Refine the plan: Update the spec based on the questions the AI asks you. Add rules for how the agent should handle unexpected inputs or failures.
-
Confirm understanding: Ask the AI to summarize the final spec. Do not move to code generation until the AI’s summary matches your exact goal.
By finishing the spec first, you give the AI a clear map. This process leads to accurate code and a reliable deployment process.
Security and Governance
AI agents must operate within strict guardrails that enforce security and governance limits. Because an agentic system acts on your behalf, applying standard data access controls is not enough—you must engineer an environment built for autonomous execution.
-
Access Control and Agent Identities: Every agent should have a unique, trackable digital identity. In Google Cloud, this means assigning a dedicated Identity and Access Management (IAM) Service Account to each agent. Apply the principle of least privilege: if an agent only needs to read specific tables to perform Conversational Analytics, restrict its IAM roles to those exact BigQuery datasets.
-
Granular Data Governance: Use Knowledge Catalog to automatically scan, classify, and tag sensitive information (such as PII). Within BigQuery, employ Row-Level and Column-Level Access Policies alongside Dynamic Data Masking. This ensures that even if an agent queries a highly sensitive table, restricted data fields are obfuscated or hidden entirely from the agent’s view and reasoning engine.
-
Safety Guardrails: Monitor task adherence and safeguard against malicious inputs. Utilize Model Armor to intercept prompt injection attacks and block policy violations before the agent ever executes an action.
Observability and Cost Control
Standard application monitoring is not sufficient for AI agents. Because they chain multiple tools and reasoning steps together, you need visibility into both their internal logic and their downstream infrastructure impact.
-
System Health and Agent Behavior: Track uptime, delay times, and error rates using Cloud Logging and Cloud Monitoring. Ensure your agent is configured to log its decision paths and tool uses directly into dedicated BigQuery audit tables, giving you a historical record of exactly why an agent chose to run a specific pipeline.
-
Quality and Safety Metrics: AI systems require continuous evaluation. Use Vertex AI Evaluation to continuously measure the agent’s accuracy, groundedness, and task completion rates. Additionally, monitor safety metrics by reviewing Model Armor dashboards for intercepted policy violations or exposed personal data.
-
Cost Metrics and Budgets: Autonomous agents can generate significant costs through rapid API calls and backend queries. To control expenses in BigQuery, monitor the BigQuery Information Schema to track the exact number of bytes billed and jobs executed by the agent’s Service Account.
Enforcing Cost Controls: Do not give an agent a blank check. Assign the agent’s specific workload to a designated BigQuery Reservation. From there, you can configure Custom Quotas—such as capping the maximum bytes billed per day per Service Account—to strictly control expenses and ensure a runaway agent loop doesn’t result in unexpected infrastructure costs.