Agentic AI only scales when the platform around it is built for production.
Everyone has an AI agent story. Most of them end the same way: a promising proof of concept that never made it to production. According to data shared during Levi9’s latest webinar, nearly 40% of all AI agentic solutions fail within the first six months. Not because the AI is bad. Because everything around the AI is.
The models are ready. The real bottleneck is the infrastructure that supports them – the data pipelines, the guardrails, the observability, the governance. And that’s exactly why platform choice matters more than most teams realize when they first start building.
To dig into this, Levi9 brought together three Lead Software Engineers – Sandra Bosic Ristic, Igor Jankovic, and Yevhen Sirotin – each with hands-on experience on one of the three major cloud AI platforms: Microsoft Azure, AWS, and Google Cloud. The conversation hosted by Dario Djurica, Principal Architect in Lev9, covered what each platform actually offers for production-ready agent deployment, and how to choose between them.
The problem isn't the LLM
A recurring theme across the entire discussion was this: teams treat AI agent projects like regular software projects, and then wonder why they fail at scale.
“The problem is not the LLMs,” Dario Djurica noted in framing the session. “The quality of the AI is good, but what is lacking is everything around the AI – the data, the integrations, the observability, the governance.”
This reframe is important for anyone planning an agentic initiative. Picking the right model is table stakes. The real engineering challenge is building the system that makes the model useful, reliable, and auditable in production. All three platforms, Azure, AWS, and Google Cloud, now offer tooling specifically designed for this. What differs is where each platform is strongest.
Each platform has a home territory
All three platforms follow a similar structural pattern: a low-code or no-code entry point for business users, and a full developer SDK for engineers who need more control.
On Azure, that means Copilot Studio for business users and Microsoft Foundry for developers. Sandra Bosic Ristic highlighted what she considers the platform’s clearest advantage: “They invest a lot in a data foundation and the connectors which your agent can use to do things or to get some data in order to work with it.”
On AWS, Igor Jankovic pointed to Agent Core Gateway as a standout feature -specifically its ability to transform existing APIs and Lambda functions into agentic tools. “Every enterprise faces the same question today: what should we do with the software that we already invested in for the previous 20 years? AWS Agent Core Gateway literally gives us that possibility to just transform these APIs into agentic tools.” The platform also offers 16 built-in evaluators that let teams benchmark agent performance before every production release.
On Google Cloud, Yevhen Sirotin made the case for Agent Search – a feature that connects agents directly to wherever your data already lives, whether that’s BigQuery, SharePoint, Drive, or an internal wiki. “You don’t build the connection, the platform does.” He also flagged something that often gets missed in cost discussions: Google provides a single unified billing invoice across runtime, memory, search, tokens, and model calls – making cost visibility considerably simpler than across fragmented surfaces.
"Start where your data is" - and mean it
The webinar’s most practical conclusion came out of the scenario comparisons. When the speakers walked through four real-world situations – business users building their own agents, teams with existing frameworks they don’t want to abandon, multi-model requirements, and cross-cloud data scenarios – the same principle kept surfacing.
“I think it’s very important to choose the platform where your data is,” Yevhen Sirotin said. Igor Jankovic agreed: “If you have data spread across all three clouds, you will probably pick the platform where your most important data lives.”
This isn’t a vague recommendation. It has concrete implications. Moving data across clouds adds latency, complexity, and cost. The managed services – guardrails, knowledge bases, agent registries – are tightly coupled to each platform’s native data layer. The more you lean on those services (which you should, because they save months of development time), the more important it becomes to start from where your data already sits.
If your users and data are on Microsoft 365, Azure leads. If you have heavy developer tooling and existing infrastructure on AWS, Agent Core gives you the best starting point. If your data lives in BigQuery or Google Workspace, the Google Cloud stack offers the most seamless integration path.
Three things to take away
1. Don't build what you can buy.
All three platforms now offer observability, guardrails, evaluation, and governance tooling out of the box. Building these from scratch with custom code or open frameworks is an option, but it likely means months of work before you ship anything to production.
2. Framework agnosticism has limits.
All three platforms claim to be framework-agnostic, and to a point they are. But the moment you start using a platform’s native managed services – memory, guardrails, knowledge bases – you’re coupling to that platform. That’s not necessarily bad, but it should be a deliberate decision made at the start of a project, not a surprise later.
3. Governance is not optional.
The EU AI Act is approaching. Dario Djurica framed it plainly during the webinar: the difference between a good AI solution and a significant regulatory fine can come down to how well your governance and compliance dimensions are covered. All three platforms offer tooling here – but you need to use it.
The platform landscape for AI agents is evolving fast. The speakers noted that the amount of change across all three platforms in just the last six to twelve months is “simply incredible” – and that conclusions drawn today may look different in a few months.
If AI agent deployment is on your roadmap, the full webinar is worth your time. And if you want to work through the platform decision for your specific context – where your data lives, what your team already knows, what your compliance requirements look like – Levi9’s team is ready to help you figure it out.
