Tech Explained: Here’s a simplified explanation of the latest technology update around Tech Explained: Tech Giants Split on How to Scale Agentic AI in Simple Termsand what it means for users..
With all the real promise and unrealistic hype around agentic AI, it’s understandable to forget a foundational truth: AI is built on data. No data, no agents. This past week has seen more evidence of exactly how tightly that connection is made.
Google says enterprise teams trying to build agentic AI run into a practical constraint: AI agents can only act as well as the data they can reach and understand, but that data is typically scattered across cloud platforms, on-premise systems and older mainframes. In the Google Cloud Blog post, Google frames the core challenge as giving Gemini and other models access to accurate, well-documented data plus the metadata that explains what the data means, where it came from and how trustworthy it is—so agents can make decisions with reliable context rather than guesswork.
To address that gap, Google highlights a partnership with Ab Initio that combines Google’s Data Cloud components (BigQuery for storage and analytics and Dataplex Universal Catalog for organizing and governing data and metadata) with Ab Initio acting as a “neutral hub” that connects hundreds of sources and standardizes metadata across a hybrid, multi-cloud environment. Google says this approach keeps data distributed while unifying metadata, adding detailed lineage (including field-level tracing) and governance features that support auditability, compliance and clearer explanations of how outputs were produced. The goal, as Google describes it, is to give Gemini richer context so agents can reason more accurately and operate with more transparency and control in large enterprises.
AWS Manages the Agents
Data is also important in managing and evaluating agents. AWS says the hardest part of building AI agents is not getting a model to “talk,” but proving the system works reliably when it has to take many steps, use tools and pull information from memory. In the AWS article, Amazon explains that traditional AI testing often looks only at the final answer. That can hide the real reasons agents fail. AWS argues teams need evaluation that treats an agent like a full system and measures what happens along the way—whether it understood the user’s request, chose the right tool, used the right inputs, retrieved the right information and completed the task.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
The data angle is central. AWS says strong evaluation starts with good inputs, especially trace files that capture what the agent actually did step by step, plus “golden” datasets with ground-truth answers for regression testing as the agent changes over time. Amazon describes a workflow where teams feed in traces, automatically generate metrics, publish results to dashboards or storage, and monitor for performance decay in production with alerts and periodic human review.
The metrics cover final response quality (such as correctness and relevance), task success, tool-use accuracy (tool choice, parameters, error rates), and memory quality (whether the agent retrieved the most relevant context). In Amazon’s own systems, AWS says teams use historical logs and simulated interactions to build datasets that let them compare agent behavior to known outcomes, so they can pinpoint failures and improve reliability before and after deployment.
Advertisement: Scroll to Continue
Golden Pipelines
Can a “golden pipeline” help? VentureBeat reports that many enterprise agentic AI projects are getting stuck on a “last-mile” data problem: companies can prepare clean, stable data for dashboards, but agents need operational data that is messy, constantly changing and needed quickly for real-time decisions. The article argues that the biggest failures often happen when imperfect data reaches real users, not when the model generates text. It highlights a vendor called Empromptu, which says the fix is to treat data preparation as part of the AI application itself, so teams spend less time manually wrangling data and more time validating whether the agent is actually working in production.
The piece describes Empromptu’s “golden pipelines” as an automated layer between raw operational data and AI features. VentureBeat says the system ingests data from many sources (files, databases, APIs and unstructured documents), cleans and structures it, fills gaps with labeling and enrichment, and applies governance controls like audit trails and access rules. A key point is the built-in evaluation loop: every data transformation is logged and tied to downstream agent performance, so the pipeline can detect when a “cleanup” step makes the agent less accurate and then surface that degradation.
VentureBeat also includes a customer example, VOW, an event platform that used the approach to turn inconsistent floor-plan and ticketing data into usable inputs for an AI-driven feature, where accuracy and speed mattered.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.