Why agentic AI needs data (and humans)

Between the recent Snowflake Summit, Databricks’ Data + AI Summit and IBM’s new watsonx.data intelligence, much of the recent conversation about enterprise AI has revolved around agents and data.

“The trends have been around gen AI, and now we all are thinking about how we harness the power of agents with humans in the loop, then building towards fully autonomous agents,” said Stephanie Valarezo, a Program Director for IBM Data & AI, in an interview with IBM Think. “And it all comes back to the data.”

Agent adoption keeps growing. For example, AI agents now create four times the number of enterprise databases that humans do on Neon, an open-source serverless database startup that helps developers build on PostgreSQL. “Every engineer is becoming an AI engineer,” said Neon CEO Nikita Shamgunov during the Data + AI Summit.

However, when it comes to deploying agents, there are still many challenges to solve. “Ultimately, why are companies building agents?” said Tim Richer, a Director of Product Marketing, Data and AI at IBM, in an interview with IBM Think. “They’re trying to have productivity. They’re trying to improve their workflows, their business domains, internal functions, sales, marketing, HR—they’re trying to reinvent in more thoughtful ways. And to build agents, you need trusted data.”

This, of course, was at the heart of the Data + AI Summit last week in San Francisco. Databricks made several announcements regarding its data and AI offerings, including the launch of Lakebase, a fully managed Postgres database integrated with Lakehouse and built for AI. (Lakebase leverages technology from Neon, which it recently acquired.)

“The industry is very focused on artificial super intelligence (...) But that’s not what our customer base wants to do,” said Ali Ghodsi, Databricks CEO, in a press meeting during the summit.

Enterprises are more focused on day-to-day applications, such as helping with HR. “Every customer we have wants to build agents, but it’s not as easy as people think,” Ghodsi added. “There’s a lot of challenges in the industry as a whole around building agents, despite the hype.”

According to Ghodsi, those challenges include concerns around evaluating or benchmarking agents and costs associated with their deployment. Accordingly, Databricks announced Agent Bricks, a product that enables businesses to generate high-performing AI agents powered by their enterprise data while also evaluating the efficiency and accuracy of their work.

For IBM, the answer to the data challenge comes with watsonx.data and watsonx Orchestrate.

“This is the central nervous system—to be able to manage not just the agents that are sitting on top of agents plural, but that might be in other parts of the data ecosystem as well,” Valarezo said. “You have to think about the entire ecosystem,” she continued. “It’s going to be across different clouds. It’s going to be in different applications. So that’s one thing that an enterprise should not lose sight of. There’s no easy button.”

Of course, enterprise adoption of agentic AI requires trustworthy models and agents that are not vulnerable to malicious actors.

“I think the whole industry is trying to get its head around cybersecurity, data security, AI risk,” said Richer. “Things like that are sort of common industry pain points that we’re all working through together. And I think IBM is particularly focused on the trust and the governance and the security and the enterprise considerations that [are] required to really scale these innovative technologies.”

One of the most important conversations happening in the agentic AI space is, of course, related to the workforce—and the role humans will play in an agentic world. The IBM Institute for Business Value (IBV) recently published a study based on a survey of executives that paints a picture of a world where people and agents work in tandem: 84% of the CEOs surveyed believe that AI agents will collaborate and transfer knowledge with humans.

“From an agent perspective, there’s going to be a lot of things that will be human in the loop, where my job might get easier—but that’s because I can tap into all this information and come to faster conclusions, or because I’m orchestrating a set of agents,” said Valarezo.

Ghodsi, for his part, is doubtful that enterprises will be able to remove humans from the loop anytime soon. “People underestimate how hard it is to completely automate a task,” he said. “Look at self-driving cars, and how hard it is to just get that last mile [to get everything perfect, NDLR]. And given that the AI occasionally gets things wrong, we want, as a society, someone to be responsible. You can’t just automate the different humans. I think it’s very far off.”