By Sary Libreros | Cloud Architect Consultant at Strata Analytics Group
An Amazon team put a language model to work on a seemingly simple task: matching author records with product listings. The final bill reached to $1.8 million, 860% over budget, according to reports the Financial Times. And the most disturbing fact is not the number: the cost overrun went undetected for five months.
The case circulated as a cost issue, but it reveals something deeper: an organization with all the resources available deploys artificial intelligence faster than it can see it, attribute its use, and limit its capabilities. And this is not an isolated case. It’s how the agentic ecosystem is growing almost everywhere.
Behind this growth lies a much older problem. An organization’s information is scattered across dozens of systems that use different names and sometimes contradict each other. An agent can have access to this entire universe and still respond incorrectly because no one has explained what each term means or how it relates to the others. For two years, we searched for the solution within the model itself—which reasoned best, which cost the least per token. The bottleneck was elsewhere.
Governing agents, in practice, begins with governing that context. This is the gap that AWS has been systematically closing, and it’s worth examining how they’re doing it.
Amazon Quick was the dress rehearsal.
When AWS launched Amazon Quick to the general public in late 2025, many saw it as just another assistant. In retrospect, it was proof that a knowledge graph could support a user’s experience by asking natural language questions about their own data.
This graph catalogs datasets, dashboards, and metadata, and learns from the user’s habits. It works, and it works in production. The investment AWS has made in Quick over the past year was not an isolated product initiative. It was building the infrastructure on which to build something bigger.
From Personal to Organizational Scale
At this year’s AWS Summit it was announced AWS Context. Context is a service that takes that same knowledge graph and scales it. Where Quick learns data from a single person, Context maintains a shared, governed graph that the entire organization can query, with relationships between systems and business rules that no single view could contain.
The service infers how a company’s data is connected: which tables connect to which, what the columns mean, and how different sources fit together. It then represents all of this as a graph that agents can traverse at runtime, combining semantic matching with relationship reasoning, rather than searching for keywords in a repository.
And this is where AI governance comes in.
This is the point we are most interested in emphasizing, because it is the one that is usually left out of the conversation.
The agent ecosystem is growing at a rate no organization was prepared to govern. Agents are multiplying, consuming data from different systems, invoking tools, making decisions, and generating costs. Without a governed context layer, each agent constructs its own version of the truth, and this fragmentation goes unnoticed until it becomes expensive to fix.
What AWS is building isn’t just an AI capability. It’s a control plane. AWS Context didn’t arrive alone: at the storage layer, Amazon S3 annotations allow you to attach context directly to each object in a mutable and queryable way, and they’re already generally available. At the catalog layer, business context and semantic search for Glue Data Catalog are in preview, allowing you to enrich tables and columns with business descriptions, glossary terms, and resources that guide agents toward query patterns and usage rules. AWS Context synthesizes these signals into the graph that the agents query.
AWS Context is still listed as coming soon, with no pricing or regions announced. However, you can already pilot S3 annotations and test catalog enrichment in regions where the preview is available. Meanwhile, those who want to start working on ontological modeling without waiting have the option to Context Ontology Accelerator, the AWS open source project that allows you to build the business ontology and serve it to agents via MCP.
The model is no longer the limit. Today, the limit is the context we are able to give it, and the difference between an impressive proof of concept and an agent that reaches production is almost always at that level.
How can we help you?
At Strata Analytics, we’ve been supporting organizations in the region on this very journey. If you’re exploring this for the first time and want to scale your AI initiatives, we can help you assess how ready your data ecosystem is to support agents in production, identify gaps before they become costly, and design a roadmap that prioritizes the right use case instead of trying to do everything at once.
If what you’re looking for is centralized and automated AI governance, we design the semantic and ontological layer, the approval and curation model, the traceability of agent decisions, and the observability of consumption and cost, on the infrastructure you already have.
Who in your organization can say today how much is being invested in AI, in which use cases, and with what return?
If two agents give your steering committee different figures about the same customer, would you know which one is correct and why?
How many of your AI initiatives went beyond proof of concept and are now generating results that the business can measure?
If any of these questions do not have a clear answer today, we invite you to talk so we can help you.
About the author
Sary Libreros is an AWS Community Builder and Cloud Architect consultant at Strata Analytics Group, a data analytics and artificial intelligence company that helps Latin American organizations develop accessible, context-aware, and culturally adaptable AI systems across the telco, finance, retail, and healthcare sectors.
