Google Cloud's Gutmans: 'We can truly build autonomous agents that really act on your behalf'

Andi Gutmans has spent the better part of three decades at the intersection of open-source software and enterprise infrastructure: co-creating PHP, building and selling Zend Technologies, running analytics at AWS, and now overseeing Google Cloud’s entire data estate as VP and GM of its Data Cloud.

Speaking to Capacity at Google Cloud Summit London earlier this month, Gutmans made a direct case: the agentic era is not a roadmap item. It is already in production, and the telecoms industry is one of its most compelling proving grounds.

What does the Agentic Data Cloud actually mean for a telecoms operator?

Gutmans is quick to ground it in operational terms. “In order to get the full advantage of autonomy and agentic capabilities, we’re working very closely with customers like Vodafone and Verizon to build digital twins of their networks, and then use agents to drive autonomous network operations,” he says. “That’s actually a very common use case we’re seeing right now.”

The appeal, he explains, lies in Google Cloud’s ability to combine data platform capabilities, graph mapping, full-text search, geospatial data, with AI and agentic orchestration. “We can bring in Google’s differentiated data sets, like mapping data and Earth Engine data, and combine that with our strength in AI and agenting.”

The data problem no one wants to talk about

Before any of that is possible, however, enterprises face a more fundamental challenge: their data estates are a mess. Gutmans cites a customer who came to Google Cloud with 20,000 database tables they needed to activate for AI agents. “Manually curating 20,000 tables, you just can’t hire enough data stewards to do that,” he says.

Google’s answer is its Knowledge Catalog, which uses agents to automatically enrich and contextualise data, identify relationships by analysing query logs, and build enough structured knowledge for other agents to operate against it. “We shortcut the path from 20,000 raw tables to actually having enough knowledge about their data state to activate agents much, much faster than if they’d had to manually curate.”

For multi-cloud environments, which, he notes, describes almost every enterprise, Google’s Borderless Lakehouse extends that activation across AWS, Azure, on-premises systems, and SaaS applications including Salesforce and ServiceNow. “The enterprise data landscape is very messy, very complex and 90% of that data is unstructured, which isn’t even catalogued today.”

Why AI projects stall at the pilot stage

The most common reason an AI project fails to reach production, according to Gutmans, is scope. “It’s trying to boil the ocean, trying to go too broad with what you’re trying to solve,” he says. His advice is consistent regardless of sector: pick a specific use case, define clear success metrics, get it into production, and build trust with users before generalising the pattern.

“Don’t think about this too generically early on in your journey, where you’re like, ‘I’m going to build a gazillion agents to run my business.’ Start with key use cases that can drive value.”

Agents critiquing agents

One of the more striking developments Gutmans describes is the emergence of agent-verification architectures, essentially, agents whose job is to quality-assure other agents. “You can have three agents vote, and if three agents agree that the answer is high quality, it’s probably a high-quality answer,” he says. “That’s just a very different way of thinking.”

A year ago, he argues, the reasoning capabilities of foundation models weren’t strong enough to make this viable. Now, Google has rewritten all of its first-party agents to take advantage. “We’re actually at the point now where you can truly build autonomous agents that really act on your behalf. Every individual contributor can have a team of agents working for them in parallel.”

Where trust still draws the line

Full autonomy remains a spectrum, not a destination, at least for now. Gutmans uses the analogy of Waymo: statistically safer than a human-driven Uber, yet still uncomfortable for many passengers. “As the stakes go up, that trust has to be built, and the technology also has to get better.” In practice, he says, operators are already comfortable with agents handling customer support escalation decisions autonomously. Large financial commitments are another matter. “If the order is £50,000, there’s no reason for an autonomous agent not to place it. If the order is £20 million, typically they would want a human in the loop.”

The direction of travel, he is clear, is towards more autonomy, not less. “That pendulum is going to swing more and more towards full autonomy.” The question for enterprises right now is not whether to start, but where. “Once you start, you suddenly realise what the art of the possible is and it’s usually way more than you actually imagine.”

Google Cloud’s Yasmeen Ahmad: The ‘dashboard fallacy’ is holding back enterprise AI agents