AI reality check: Here's what three companies learned building wallets, homes, and games

Three very different companies stood on the same stage at Google Cloud Next last week and said roughly the same thing: the hard part of AI agents isn't launching them, it's living with them afterward. Citi, Home Depot, and Capcom each described agents that have moved well past demo status into daily operations touching real customer money, real store traffic, and real game production schedules. And the lessons they've learned sound less like hype and more like the unglamorous work of building infrastructure that doesn't fall over.

Citi's version is called Citi Sky, a client-facing voice and video agent that Andy Sieg, the bank's head of wealth, compared to the arrival of the ATM. The pitch is straightforward: Citi manages roughly $1 trillion for its wealth clients, but those same clients hold another $5 trillion elsewhere, and hiring enough humans to chase that money would mean adding thousands of advisors. Sieg says Citi Sky gives the bank effectively unlimited capacity instead, including the ability to blast simultaneous portfolio reviews to every client the moment something moves markets in Washington. But the governance angle is arguably the more interesting bit. Citi didn't just bolt Google's models onto its business; it layered its own compliance and wealth-management rules on top and treats the agent like a licensed employee, subject to the same securities rules as a human advisor. That also means every interaction is now auditable in a way human conversations, which vary advisor to advisor and often go unrecorded, never were.

Home Depot's approach is quieter but arguably more structurally ambitious. CIO Angie Brown described a single underlying system, built on BigQuery data with Gemini Enterprise and Google's ADK framework, that powers the company's Magic Apron assistant across the website, in-store pilots, and a new phone system that replaces the old IVR menu entirely. The point is consistency: ask about deck materials online and you get the same answer calling the store or asking Magic Apron on the professional contractor channel, which now helps price jobs. Brown says that portability is already showing up in the numbers, with shopping-agent users converting at a higher rate and the new phone system resolving calls about four times faster than before.

Capcom's story is the most concrete on raw output. Shinichi Inoue, who oversees the company's game development platform and AI solutions, said AI agents are now saving roughly 30,000 hours per month per project, largely by taking over the grinding regression testing that eats enormous chunks of time on modern titles — the endless cycles of watching, listening to, and playing through builds looking for bugs. With about three-quarters of Capcom's 3,700 employees in development roles, that's a meaningful chunk of labor reassigned. Inoue was careful to frame the win less as a cost-cutting story and more as a creative one: freeing testers and designers from repetitive verification work, he argued, is what actually lets Capcom's people focus on the parts of game-making that require imagination rather than checklist discipline.

What ties these three together isn't the technology stack, since all three happen to run on Google Cloud, but the discipline around it. None of these companies just dropped a chatbot on customers and called it done. Citi built compliance scaffolding, Home Depot built one unified backend across every channel, and Capcom pointed the saved hours specifically at creative bandwidth rather than headcount reduction. That's a more boring story than the usual agentic-AI pitch, but boring is often what shipping to millions of real customers actually looks like.