What underpins AI—data, databases, application stacks, orchestration patterns, and more—brings a sobering reality to the otherwise thrilling promise of AI technology. However, developing a strong understanding of how this interplay works, as well as the evolving trends designed to fuel AI, is crucial for real-world applications.
Sailesh Krishnamurthy, vice president of engineering, Google Cloud, and Joe Zimmermann, staff program manager, Google Cloud Databases, joined DBTA’s webinar, Beyond the Hype: Real-World Applications of Gen AI and Databases, to explore real-world use cases and valuable insights into today’s top technologies and patterns shaping the business landscape.
“The world is changing about once every other week,” joked Krishnamurthy, as he pointed to various sentiments pervading the tech landscape: “Everyone’s ‘vibe coding’ now,” “SaaS is dead,” and “Everything’s an agent.”
Deciphering what’s hype and what’s real will be a crucial skill as technology continues to evolve and transform at a dramatic pace.
As AI relates to databases, Krishnamurthy looked at internal versus external workload, highlighting how these areas are differentiated by their need to shrink or expand knowledge. Internal workloads need to be able to shrink their data estates to add context and nuance, while external workloads need the ability to expand the tiny slice of what they can see by contextualizing with other unstructured data.
While the order in which database processes and search functions differ for each workload, one pattern connects them: Both workloads need the ability to combine structured and unstructured data with flexible and secure agentic techniques.
Enter agentic AI, an innovation victim to massive hype this year. To help shed these rose-colored glasses, Krishnamurthy broke down exactly what an agent is:
- An application that uses generative AI (GenAI) models to think and act towards goals
- Automate complex enterprise workflows
- Use reasoning and planning to determine steps
- Utilize tools to access data and take actions, such as querying databases or calling APIs
Crucial to agent function is the use of tools, a word which Krishnamurthy noted may be “even more abused than agents.” Tools are essentially function calls that allow the large language model (LLM) to act as the engine for an AI agent to take actions beyond simply generating text.
When an enterprise has many tools, however, challenges arise, such as a lack of understanding of tool interdependencies creating unique security risks that must be addressed at the individual tool level. Google’s MCP Toolbox for Databases resolves these issues, acting as an open source MCP server that allows enterprises to develop tools easier, faster, and more securely by alleviating various burdens from the process, such as connection pooling, authentication, and more.
It’s undoubted that the agentic AI wave has seriously disrupted many industries, unlocking significant efficiencies through its automation capabilities and its multimodal nature. Through this disruption a new application stack has emerged, explained Krishnamurthy, where the combination of agentic, search, multi-modal support, chat, and natural language interface can generate brand new user experiences. This stack empowers developers to securely leverage AI on data, allowing them to access data directly and therefore bypassing complex app tiers to intercept user intent while adhering to security controls.
However, what makes a database AI-ready to support this new stack? According to Krishnamurthy, it’s “not just a database with one or two new features, it’s a collection of deep capabilities that actually should make us reconsider what the boundaries of the database are.” He further highlighted the following as top characteristics for an AI-ready database:
- Semantic Context and Programmatic Controls: AI-ready databases need extra contextual information to be made explicit and accessible to AI and must enforce additional controls and constraints.
- Interactive Interfaces: In the absence of conventional applications with prescribed logic, agents and databases must jointly and interactively collaborate to read and write the correct data to meet the user’s needs, blurring the boundary between application and database.
- Non-deterministic: Semantic operators and the increasing use of unstructured data and natural language mean that database operations are increasingly non-deterministic. Probabilistic accuracy and recall increasingly displace strict definitions of correctness.
Krishnamurthy and Zimmermann then led webinar viewers through a variety of Google Cloud innovations poised to deliver the AI-ready database, such as AlloyDB, the industry-leading PostgreSQL for operational, analytical, AI apps. AlloyDB makes it easier to build intelligent applications and agents, featuring a plethora of capabilities that power the new application stack.
This is only a snippet of the full Beyond the Hype: Real-World Applications of Gen AI and Databases webinar. For the full webinar, featuring more detailed explanations, a demo of AlloyDB, and more, you can view an archived version of the webinar here.