LONDON — The era of AI experimentation is over, and businesses should prepare for the production-scale deployment of agentic AI.
That was the cloud giant’s overriding message at its Google Cloud London Summit conference this week.
“The agentic enterprise is happening right here, right now,” Maureen Costello, vice president for the U.K., Ireland and sub-Saharan Africa at Google Cloud, said during her keynote.
Recent Google product launches have joined this agentic momentum, including the Gemini Enterprise Agent Platform, an environment for deploying and managing AI agents at scale.
Costello also showcased Gemini Omni, a multimodal model designed for low-latency applications, and Gemini 3.5 Flash, which the vendor described as its strongest model yet for agentic workloads.
With its recent product rollouts, the tech giant is positioning itself as a one-stop shop offering users the full stack required to build, deploy and govern enterprise AI — which the vendor also bills as the “front door” to a range of AI applications.
Other recent releases include Knowledge Catalog, an enterprise context layer designed to connect AI systems with organizational data, and Google AI Threat Defense, a security offering aimed at helping organizations identify and mitigate emerging AI-powered incursions.
Beyond the product releases, much of the conversation at the conference turned to what embedding AI throughout an organization actually demands, particularly in governance, transparency and accountability.
From Pilots to Production
For enterprises deploying Google’s agentic tools at scale, the shift from experimentation to real-world deployment is already well underway.
A roundtable featuring leaders from THG Ingenuity, Kingfisher, Rightmove and Deloitte offered a ground-level view of what that transition actually looks like and what it requires from organizations.
Jo Drake, CTO of e-commerce vendor THG Ingenuity, described the company’s AI shopping assistant built on the Gemini Enterprise Agent Platform as a case study in what proactive, personalized AI can unlock at scale.
The returns, Drake argued, come from AI’s ability to replicate the inherently conversational nature of shopping.
“Even before e-commerce, you would walk into a shop and discuss what you’re looking for with a person,” she said. “We’ve used Gemini to really bring that conversation to life.”
In addition, conversations between chatbots and customers are themselves proving a rich source of insight into other consumer demands.
“This insight has become really, really useful to brands in how they actually market and build their brand online,” Drake said.
Culture Before Code
“Moving from a chat interface to a world of autonomous agents is a completely new paradigm,” Hayley McKelvey, chief AI officer for Deloitte tax and legal U.K., said. “It’s only when you get your hands on keyboards that you can see the art of what’s possible with agents. The scales fall from your eyes, and you realize what can actually be done here.”
Addressing workforce acceptance of AI and integrating it effectively throughout an organization is crucial, she continued.
“Do people in your organization run toward what scares them or do they run away from it?” she said. “If organizations and leaders can really tap into that highly emotional reaction — how people are really, really feeling about this stuff — that’s where progress becomes really accelerated.”
On the question of ROI, McKelvey said the industry’s current focus on tokenomics — or understanding and controlling the cost of model usage — is necessary, but risks becoming a distraction.
“Just because you can see that you spent £100 million [$132.4 million] and you budgeted to spend £100 million doesn’t mean you’ve realized value,” she said, warning that organizations risk measuring new technology against old benchmarks. “We need to stop measuring new stuff and new organizations by reference to old measurements.”
The Governance Gap
While the conference’s tone was broadly optimistic, one of its sharper undercurrents concerned what happens when AI deployment outpaces the structures meant to oversee it.
Sailesh Krishnamurthy, vice president of databases and engineering at Google Cloud, told AI Business there has been an explosion of applications built not by software engineers, but by citizen developers. This refers to non-technical users who can spin up tools and workflows using AI without writing a single line of code.
The increased accessibility of AI tools is a double-edged sword — bringing greater opportunities for innovation and development but also heightening the risk of app development and AI usage going unchecked.
The governance challenge compounds, Krishnamurthy said, when AI agents begin executing complex, multi-system workflows on the fly. Many of these bypass formal development processes, leaving no record for developers to trace.
“Suppose someone leaves, and someone else comes and says, ‘I want to know how many such things happened. I want to change it, I want to go back and remove one of these things,’ you can’t if you didn’t keep a log,” he said.
Krishnamurthy said the answer lies in consistent logs of AI actions, or “agent trajectories.” Without them, he said, enterprises are operating blind.
“You can’t govern what you don’t store,” he said. “I think a lot more of that is happening than people fully realize. The horse has already left the stable.”
The path forward, he suggested, requires treating the harnesses and evaluation frameworks that shape AI behavior as genuine engineering problems, worthy of the same investment as the applications themselves.
Despite some uncertainty, Krishnamurthy is optimistic.
“Things that are hard to do today will become easier,” he said. “The tricky thing is to do it well and do it safely. I think that’s where there are technical engineering problems, but these are being looked at, and on the consumption side, it’s quite a time to be alive.”
