Google Cloud is taking aim at improving agentic AI accuracy with a series of new offerings in its Agentic Data Cloud.
New Knowledge Catalog capabilities are meant to boost the agentic AI context layer's access to shared, defined, consistent business logic so they can act on trusted, verified information. They include the preview of AII-powered self-learning enrichment, which learns from users' behavior to mine unstructured data sources for contextually appropriate data, and the preview of a new studio experience that provides a conversational interface for viewing an organization's entire data estate and exploring context.
Meanwhile, Google's borderless lakehouse, which is built on Apache Iceberg and enables users to activate data regardless of whether it lives in the cloud, on-premises or in SaaS platforms, now features cross-cloud interconnect to enable users to query records that live on AWS and Azure using natural language.
In addition, it now includes a direct integration with the Knowledge Catalog to connect agentic AI tools with context, Google's Spanner Omni database to extend multi-model capabilities to on premises, edge and multi-cloud environments, and an integration with AlloyDB to directly query Iceberg tables in cloud storage without requiring pipelines to move data.
"This extends what Google announced when it introduced the Agentic Data Cloud," Donald Farmer, founder and principal of TreeHive Strategy, told TechTarget. "The big change is that Google no longer expects customers to bring their data into BigQuery but now wants to be where agents get governed access and business context for data that stays in AWS, Azure, Snowflake, Databricks, SAP or Salesforce."
Michael Ni, an analyst at Constellation Research, likewise noted that the new features are significant because they enable agentic AI tools to work across different environments. In addition, he pointed out that with its Knowledge Catalog, the hyperscaler is making business context a shared infrastructure.
However, Google is not alone in doing so.
"Microsoft, Databricks, Snowflake and AWS are all bringing data, business context and AI agents closer together. … The individual capabilities aren't all unique, but bringing them together could reduce a lot of the integration and operating work customers face," he told TechTarget. "Microsoft is probably the closest overall competitor, with Databricks particularly strong on the data and context side."
Connecting Agents to Context
Google's latest capabilities build on those unveiled in April when the tech giant first unveiled its Agentic Data Cloud. A combination of factors, including customer feedback, gave Google the impetus for adding the new features, according to Sirish Chandrasekaran, vice president of strategy and growth for its Data Cloud.
"The impetus here is the same as any product development at Google Cloud, which is a combination of watching emerging technology trends, a clear understanding of what customers actually need to run those technologies in production, and of course, our internal engineering teams' constant innovation and ability to solve these bottlenecks from the inside out," he said.
The big change is that Google no longer expects customers to bring their data into BigQuery but now wants to be where agents get governed access and business context for data that stays in AWS, Azure, Snowflake, Databricks, SAP or Salesforce.
Donald Farmer
Founder and principal, TreeHive Strategy
Knowledge Catalog is perhaps the most valuable of the new Google features aimed at aiding customers connect agentic AI with context, according to Ni.
In benchmark testing conducted by Google but not verified by an independent third party, developing and deploying agents that draw on its Knowledge Catalog makes agents nearly four times more accurate than agentic AI tools not integrated with the Knowledge Catalog. In addition, the Knowledge Catalog was shown to help agents reduce token usage by up to 75% and lower latency by up to 33%.
"Knowledge Catalog stands out because Google is moving beyond cataloging data to managing the business context agents need to do their jobs," Ni said.
He noted that enterprises have spent years building semantic layers that define metrics and business rules across their systems. Knowledge Catalog enables users to build on top of those semantic layers in platforms such as Snowflake, Databricks, SAP and Looker by using AI to add operational knowledge stored in documents and workflows.
The result is business context for agents that is consistent across entire organizations.
"Google is now adding automated ways to collect knowledge, along with tools for people to review, manage, and share it across agents," Ni said. "The opportunity is to maintain that business understanding once rather than having every team rebuild it for each new agent.
Farmer, meanwhile, called out the significance of AlloyDB being able to read Iceberg tables without requiring extract, transform and load pipelines. In addition, he noted how valuable it is to ensure that users can get fixed, verified answers to repeated questions because saved operational reports can run as queries are valuable.
"It matters because many -- if not most -- business questions don't need a language model each time they are asked, just a lookup to a report," Farmer said.
The missing links
Google's latest features aid the tech giant's push toward providing a platform that enables customers to connect agents and other AI tools with relevant context, according to Farmer.
However, it does so unevenly, he continued, noting that transparency related to how agents arrive at their responses and the cost to do so is missing.
"Google needs to be able to show, for each task, what the agent used, whether the answer was right, and what it cost," Farmer said. "They … offer no way to see the cost of a completed task across BigQuery, AlloyDB and model use. Without that, 'trusted' and 'low cost' are big claims but customers cannot readily check them."
In addition, he advised Google to stop competing with the Open Sematic Interchange by making Knowledge Catalog available as an open service -- with Google's policies and pricing applied -- that any agent, on any platform, can use.
"It would give Google a way to attract organizations whose data and agents are elsewhere, which is the vast majority of cases," Farmer said. "Google has already accepted other vendors' semantic models as inputs, so offering its context layer as an output to other vendors' agents is a consistent next step."
Ni likewise noted that Google is effectively building an infrastructure for customers to scale agentic AI. However, more needs to be done to confidently empower agents to autonomously execute complex processes rather than respond to simple queries.
"There's a difference between an agent that can answer a question and one that can safely run a business process," Ni said.
For example, he pointed out that a customer service agent might understand a customer's history and be able to recommend a refund when one is requested. But whether the agent can be trusted to determine if a refund is authorized, complete the transaction and recover if something goes wrong is yet to be proven.
"That's the next level of enterprise AI, and it's where the value moves from productivity to operating results," Ni said.
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.
