Vanguard built a Virtual Analyst on AWS around eight principles of AI-ready data

Vanguard, one of the world's largest asset managers with a portfolio of more than $9 trillion, has published a detailed breakdown of how its team built the Virtual Analyst, an internal AI tool on AWS that answers business teams' analytical questions in natural language, without any code.

Data first, models second

The key lesson from Vanguard's experience: AI transformation starts not with choosing a neural network or buying compute, but with putting data in order. The company's engineers formulated eight principles of AI-ready data, which formed the foundation of the whole project before the first line of model code was written.

The principles cover the full lifecycle, from semantics and structure to security and monitoring:

• Unified taxonomy: a shared vocabulary for all metrics, KPIs and business entities, so that "yield" in one division means the same as in another
• Data lineage: traceability of every metric from the primary source to the analytical data mart
• Timeliness: a guarantee that data is up to date at the moment of each query
• Machine-readable metadata: schemas and descriptions that are understandable not only to people but also to automated systems
• Access control: granular security policies at the row and column level
• Quality monitoring: automated, real-time checks of data correctness
• Standardized formats: unified schemas and conventions adopted by all teams
• Documentation: reproducibility of every dataset and explainability of every calculation

Without these principles, AI models hallucinate on incorrect or ambiguous data. Vanguard tackled the problem systematically rather than piecemeal, and that became the foundation of its success.

AWS under the hood

For the technical implementation, Vanguard used a combination of AWS services. Amazon S3 serves as a single data lake that consolidates sources from different divisions. AWS Glue Data Catalog handles ETL pipelines and centralized metadata storage, which is where the descriptions, schemas and business definitions of all datasets live.

Model training and deployment run on Amazon SageMaker. Complex multi-step processes are orchestrated through AWS Step Functions, while data quality and pipeline performance are monitored through Amazon CloudWatch, with alerts and dashboards configured for the team.

The Virtual Analyst runs on top of this infrastructure: it accepts questions in natural language, translates them into data queries and returns structured analytics with charts and text interpretations. Business teams get insights without SQL, Python or a queue of requests to data specialists.

Business results

"The path to AI-ready data is not a one-off project but an operational culture," the authors of the case study emphasize on the AWS

Machine Learning blog.

Vanguard reports a concrete, measurable effect. The time needed to prepare standard analytical reports has fallen several-fold. Portfolio managers and risk analysts put questions directly to the system, instead of assigning tasks to data analysts and waiting hours or days for an answer.

Importantly, the Virtual Analyst does not replace human expertise. It takes over the routine part of the work, namely aggregating, filtering and initially interpreting data, and frees analysts for higher-level tasks: developing hypotheses and making strategic decisions.

What this means

The Vanguard case is one of the most detailed public descriptions of how a large financial company builds AI analytics in production. The eight AI-ready data principles are a practical checklist for any organization that wants real business impact from AI, rather than an attractive pilot on perfectly prepared test data.