Jul, 134 min to read
Automation
News · Google's Cloud Next '26 customer roundup, read for what's actually shipping
Apr, 224 min to read
AutomationGoogle's Cloud Next '26 customer roundup, read for what's actually shipping
Behind ten marquee agent deployments is one recurring platform and a handful of load-bearing numbers
Gemini Enterprise is the thread, not the ten logos
The post presents ten independent stories, but most of them run through the same product. Home Depot's Magic Apron and its phone agent are built on Gemini Enterprise and Gemini Enterprise for Customer Experience. Merck's up-to-$1-billion platform deploys Gemini Enterprise across R&D, manufacturing, commercial and corporate functions. Mars named Gemini Enterprise its "primary AI operating system." Unilever's procurement solution uses the Gemini Enterprise Agent Platform, as does Vodafone's AI Concierge. Virgin Voyages' Rovey is powered by Gemini Enterprise plus Google Distributed Cloud.
That repetition is the real announcement. Google is not showcasing ten bespoke research collaborations; it is demonstrating that one commercial platform is being adopted as a standard layer inside large enterprises. The customer names supply credibility; the product name supplies the actual product story.
Which numbers are doing work, and which are decoration
A few figures in the post are concrete and testable. Capcom's playtesting agents log more than 30,000 hours of testing per month, navigating game worlds to find bugs, visual glitches and audio inconsistencies. Home Depot reports its phone agent could identify a caller's need within 10 seconds in a nationwide pilot. Citadel Securities cites TPUs running AI workloads up to four times faster at 30% lower cost, collapsing work that took days into minutes. Tata Steel deployed more than 300 specialized agents in nine months. Virgin Voyages says Distributed Cloud cut production timelines by up to 60%.
Other claims are aspirational rather than measured. Merck's investment is "valued at up to $1 billion" and aims to boost productivity for 75,000 employees, but no outcome is reported. Unilever's "personalization at scale" for 3.7 billion people describes ambition, not a deployed result. The reader's job is to separate the pilot metrics from the promises — the strongest evidence here comes from Capcom, Citadel and Home Depot, where a specific number attaches to a specific task.
The infrastructure underneath the agents
Two of the ten stories are really about hardware and edge deployment, not agents per se. Citadel Securities' quote centers entirely on TPU Ironwood and running thousands of parallel chips for a single workload — an argument that researchers should be limited by their own creativity rather than platform scale or economics. Virgin Voyages' Rovey depends on Google Distributed Cloud to keep the concierge agent running on ships with limited connectivity on the open water.
These inclusions matter because they show where the constraint actually sits. An agent is only as useful as the compute it can access cheaply and the connectivity it can survive without. Citadel's cost-per-experiment and Virgin's offline resilience are the enabling conditions that make the agent layer viable, and Google is quietly foregrounding its silicon and edge stack alongside its models.
The shift this roundup documents: from general assistants to task-scoped fleets
The most instructive detail is how narrow most of these agents are. Capcom describes visual inspection, predictive and institutional knowledge agents — three specialized roles, not one general assistant. Tata Steel talks about a fleet of over 300 agents and a low-code platform, Zen AI, that lets non-data-scientists build and deploy their own. Unilever built a multi-agentic solution scoped to procurement decisions. Even the customer-facing tools have tight jobs: Home Depot's phone agent routes intent; Citi Sky answers "Am I financially OK?" and explicitly does not replace advisors.
That is the automation pattern worth taking from this post. The deployments that carry real numbers are the ones aimed at a bounded, repetitive task — playtesting hours, call routing, research iterations — rather than open-ended reasoning. For teams building their own agents, the takeaway is not that agents can do everything, but that the ones already producing measurable results are the ones given a single well-defined job and the infrastructure to do it at volume.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.