Share with your CTO
On, the Swiss premium sportswear brand, bet its next growth phase on collapsing the traditional boundaries between compute, data, and AI by consolidating its entire technology stack on Google Cloud and running the migration itself through coordinated AI agents. A small internal engineering team moved 24 core microservices in roughly two weeks per service, down from three months, with under five minutes of planned downtime each. An outside integrator had quoted $500,000 for a partial migration. On’s team did more, faster, without the vendor.
What this means for your business
The telling detail here isn’t the speed. It’s the $500,000 quote that never got paid. If a coordinated set of domain-specific AI agents, each handling a discrete slice of the migration workflow like codebase analysis, configuration generation, and environment validation, can displace a systems integrator on a project of this complexity, then the cost structure of enterprise infrastructure work has shifted in a way that most technology budgets haven’t yet priced in. CTOs who still reflexively scope large migration programs as SI-led engagements should ask whether that default is now an expensive habit rather than a necessity.
The architecture On used deserves close reading. This wasn’t a single AI assistant generating scripts for engineers to clean up. It was a multi-agent system where specialized agents handled distinct task domains, humans reviewed outputs and authorized production cutovers, and no consequential change executed without sign-off. That bounded execution model, where agents have defined scope and humans hold veto power over irreversible actions, is the design pattern that makes agentic workflows safe enough to run on production infrastructure. Organizations experimenting with AI agents on lower-stakes internal tooling should understand that this pattern scales, but only if the human-in-the-loop controls are built in from the start, not bolted on after something breaks.
Google Cloud’s incentive here is obvious, publishing a press release about a customer win naturally flatters the platform’s capabilities, which makes the $500,000 displaced-vendor figure worth scrutinizing rather than accepting as a clean data point. But even discounting for promotional framing, the directional claim holds. The recurring failure mode in large cloud migrations has been runaway SI costs and timeline slippage caused by the sheer mechanical volume of configuration and translation work. If agents can absorb that mechanical load reliably, the economics of the migration category change whether Google Cloud captures the benefit or not. The vendor to pressure first isn’t your cloud provider. It’s the integrator on your next statement of work.
Concept deep-dive: Multi-agent architecture
A multi-agent system splits a complex workflow across several AI agents, each trained or prompted for a narrow task, the way an assembly line assigns specific stations rather than asking one worker to build the whole product. In On’s migration, separate agents handled code analysis, infrastructure scripting, and testing validation before any human reviewed the output. The business value is that narrow scope makes each agent’s outputs auditable and its failures contained, which is what makes the approach viable on production systems rather than just internal prototypes.
Based on reporting from On Establishes Google Cloud as Enterprise AI Backbone, Beginning with Agent-Led Cloud Migration, originally published 2026-10-08 08:07:00.
