AI-Ready Data Infrastructure for Conversational Financial Analytics

vanguard · Finance · 2026 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Conversational financial analytics

Vanguard wanted business users and analysts to query complex financial data conversationally, but reliable natural-language analytics required governed, AI-ready data infrastructure rather than only a better foundation model.

Constraints:

financial analysts previously needed SQL expertise and data team support
requests could take several days to fulfill
sensitive financial data required security, governance, and authorization controls
generated queries needed business semantics as well as technical schema metadata
semantic definitions and exemplars had to evolve without breaking KPIs or SLAs

Architecture

Vanguard built a Virtual Analyst around AWS data and AI services: Amazon Bedrock for foundation-model natural language understanding, Bedrock Guardrails for protected inputs and outputs, ECS for scalable compute, DynamoDB for conversation persistence, S3 for storage, SageMaker for experimentation, Redshift for the warehouse, and Glue for cataloging and ETL. The core implementation work was the AI-ready data layer: ownership models, governance, metadata/catalog APIs, a semantic layer, question-to-SQL exemplars, data quality checks, versioned semantic definitions, and continuous evaluation loops.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  business-user-interface["business-user-interface"] --> virtual-analyst-application["virtual-analyst-application"]
  virtual-analyst-application["virtual-analyst-application"] --> amazon-bedrock-models["amazon-bedrock-models"]
  virtual-analyst-application["virtual-analyst-application"] --> bedrock-guardrails["bedrock-guardrails"]
  virtual-analyst-application["virtual-analyst-application"] --> dynamodb-conversation-store["dynamodb-conversation-store"]
  metadata-catalog-control-plane["metadata-catalog-control-plane"] --> semantic-layer["semantic-layer"]
  glue-catalog-etl["glue-catalog-etl"] --> metadata-catalog-control-plane["metadata-catalog-control-plane"]
  glue-catalog-etl["glue-catalog-etl"] --> redshift-warehouse["redshift-warehouse"]
  semantic-layer["semantic-layer"] --> virtual-analyst-application["virtual-analyst-application"]
  semantic-layer["semantic-layer"] --> redshift-warehouse["redshift-warehouse"]
  question-sql-exemplar-library["question-sql-exemplar-library"] --> amazon-bedrock-models["amazon-bedrock-models"]
  question-sql-exemplar-library["question-sql-exemplar-library"] --> continuous-evaluation-suite["continuous-evaluation-suite"]
  continuous-evaluation-suite["continuous-evaluation-suite"] --> virtual-analyst-application["virtual-analyst-application"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`business-user-interface`	Conversational entry point where analysts and stakeholders ask financial data questions in natural language.	`virtual-analyst-application`
`virtual-analyst-application`	Application service running on scalable compute that coordinates language understanding, guardrails, conversation state, and data access.	`business-user-interface`, `amazon-bedrock-models`, `bedrock-guardrails`, `dynamodb-conversation-store`, `semantic-layer`, `redshift-warehouse`
`amazon-bedrock-models`	Foundation models that interpret user questions and help generate analytical SQL.	`virtual-analyst-application`, `semantic-layer`, `question-sql-exemplar-library`
`bedrock-guardrails`	Controls AI inputs and outputs to protect sensitive financial data.	`virtual-analyst-application`
`metadata-catalog-control-plane`	Unifies technical metadata, business definitions, lineage, synonyms, ownership, and usage context through APIs.	`semantic-layer`, `glue-catalog-etl`, `amazon-bedrock-models`
`semantic-layer`	Translates business definitions and relationships into executable query logic for SQL generation.	`metadata-catalog-control-plane`, `virtual-analyst-application`, `redshift-warehouse`
`redshift-warehouse`	Central analytical warehouse queried by the Virtual Analyst.	`virtual-analyst-application`, `glue-catalog-etl`, `semantic-layer`
`glue-catalog-etl`	Catalogs data and powers ETL jobs that consolidate accurate data.	`metadata-catalog-control-plane`, `redshift-warehouse`
`question-sql-exemplar-library`	Stores question-to-SQL examples used for few-shot prompting, benchmarks, and regression testing.	`amazon-bedrock-models`, `continuous-evaluation-suite`
`continuous-evaluation-suite`	Monitors regression, feedback, model degradation, and business-impact metrics.	`question-sql-exemplar-library`, `virtual-analyst-application`
`dynamodb-conversation-store`	Persists conversation state across the horizontally scalable architecture.	`virtual-analyst-application`

Data flow

Business users ask financial analytics questions in natural language. The Virtual Analyst uses Bedrock-backed language understanding inside an AWS application stack, applies guardrails to protect sensitive financial data, relies on metadata catalog and semantic-layer definitions to translate business terms into structured SQL over Redshift data, persists conversation state in DynamoDB, and checks outputs against exemplar/evaluation suites and data quality monitors so the system can improve without losing business governance.

"* Amazon Bedrock for foundation models that power natural language understanding

Amazon Bedrock Guardrails to secure AI inputs and outputs to protect Vanguard’s sensitive financial data
[Amazon Elastic Contain"

— article, stated "* Amazon Bedrock for foundation models that power natural language understanding

Amazon Bedrock Guardrails to secure AI inputs and outputs to protect Vanguard’s sensitive financial data
[Amazon Elastic Contain"

— article, stated "dshift](https://aws.amazon.com/redshift/) for centralized data warehousing

AWS Glue for data cataloging while powering numerous extract, transform, load (ETL) jobs to consolidate accurate data"

“With this layer in place, business analysts can express their understanding of data relationships in natural language that can be interpreted and translated into structured SQL queries.”article, stated

Why this architecture

Vanguard chose an AWS-integrated architecture because financial analytics needed managed AI services, data warehouse/catalog services, security controls, and enterprise governance together. The case explicitly frames the breakthrough as data architecture: metadata, semantic definitions, exemplars, change control, and evaluation make generated SQL reliable enough for business users, while Bedrock and guardrails provide the natural-language and safety layer.

4 supporting quotes

“building effective conversational AI wasn’t a machine learning challenge—it was a data architecture challenge.”article, stated

“Vanguard chose AWS for its comprehensive suite of integrated services.”article, stated

“the security and compliance features of AWS met the stringent requirements of the financial services industry.”article, stated

“A single catalog brings these two metadata types together, enabling AI systems to generate accurate queries that align with both technical structure and business meaning.”article, stated

Models used: Amazon Bedrock · Amazon Bedrock Guardrails · Amazon Redshift · AWS Glue · Amazon DynamoDB · Amazon ECS · Amazon SageMaker

Scale

Vanguard is a global investment management firm serving individual investors, institutions, and financial professionals.
“Vanguard is a global investment management firm, offering a broad selection of investments, advice, retirement services, and insights to individual investors, institutions, and financial professionals.”article, stated
Typical analytics requests previously took several days.
“The existing workflow required SQL expertise and data team support, with typical requests taking several days to fulfill.”article, stated
The exemplar library contained over 50 question-to-SQL examples.
“Vanguard built a collection of over 50 exemplars that serve three purposes”article, stated
The guidance starts semantic-layer documentation from the top 20 most-used metrics.
“Work with business stakeholders to document the top 20 metrics your organization uses most frequently”article, stated
The project reduced time-to-insight from days to minutes.
“Reduced time-to-insight from days to minutes for complex financial queries with the use of the Virtual Analyst”article, stated

Alternatives considered (roads not taken)

Treat conversational analytics primarily as a foundation-model selection problem

Rejected because: The team found reliability depended on AI-ready data foundations, semantic context, metadata management, and governance.

“This realization led to a fundamental shift in approach: instead of focusing solely on AI capabilities, Vanguard needed to build what they termed AI-ready data.”article, stated

Rely on technical metadata alone

Rejected because: Technical metadata without integrated business context creates mismatches between schemas and business questions.

“Organizations often maintain complete technical metadata but lack integrated business context, creating misalignment between technical implementations and business requirements.”article, stated

Let semantic definitions and examples change informally

Rejected because: Vanguard treats semantic definitions, exemplars, and configurations as versioned code with CI/CD, approvals, and stakeholder sign-off for KPI/SLA changes.

“Treat your semantic definitions, exemplars, and configurations as code under version control.”article, stated

“This approach requires stakeholder sign-off for changes that affect KPIs or SLAs”article, stated

Patterns

`guardrails-io-filtering` (safety)

Bedrock Guardrails protect AI inputs and outputs around sensitive financial data.

"* Amazon Bedrock for foundation models that power natural language understanding

Amazon Bedrock Guardrails to secure AI inputs and outputs to protect Vanguard’s sensitive financial data
[Amazon Elastic Contain"

— article, stated

`structured-output-contract` (data-contract)

The semantic layer translates business terms and relationships into executable SQL logic so natural-language questions map to governed query structures.

“The semantic layer operationalizes the business metadata defined in your catalog by transforming complex data structures into user-friendly formats.”article, stated

`eval-driven-development` (evaluation)

Question-to-SQL examples serve simultaneously as few-shot prompts, evaluation benchmarks, and regression tests.

“serve three purposes: few-shot prompts for the AI model (providing example question-answer pairs to guide the model’s responses), evaluation benchmarks”article, stated

`data-quality-loop` (data-quality)

Automated distributional, referential, reconciliation, and freshness checks monitor the data reliability behind generated analytics.

“Vanguard set up observability tools to monitor data reliability through automated checks”article, stated

`prompt-versioning` (change-control)

Semantic definitions, exemplars, and configurations are treated like version-controlled code with review and gated deployment.

“Change control and continuous integration and deployment (CI/CD) processes treat semantic definitions, exemplars, and pipeline configurations as code”article, stated

Tradeoffs

Invest in data architecture before scaling model use

Chose AI-ready data foundations over model-centric implementation.

Rationale: Reliable conversational analytics depended more on governed metadata, semantic definitions, and quality loops than on choosing a better model alone.

“This realization led to a fundamental shift in approach: instead of focusing solely on AI capabilities”article, stated

Use a semantic layer for business definitions

Chose standardized executable business logic over letting generated SQL infer metrics directly from raw warehouse structures.

Rationale: The semantic layer enforces shared definitions and relationships, which reduces errors and makes generated SQL align with business meaning.

“By enforcing the business definitions and relationships documented in your metadata catalog, the semantic layer enhances consistency across queries”article, stated

Begin metadata work with the most-used data and metrics

Chose prioritized rollout over documenting every asset before launch.

Rationale: The article recommends starting with frequently accessed datasets and top metrics, then expanding based on usage and feedback.

“Start with your most frequently accessed datasets and systematically document their technical metadata before expanding to other data sources.”article, stated

Gotchas (what bit them)

Foundation models alone were not enough for trustworthy financial analytics.

Root cause: Generated SQL needs semantic definitions, ownership, metadata, and quality standards that the model can rely on.

Resolution: Shift the work toward AI-ready data infrastructure with operating models, catalog metadata, semantic layers, quality checks, and evaluation suites.

“The most sophisticated foundation models require proper data foundations to deliver reliable results.”article, stated

“without clear ownership models, semantic definitions, and quality standards that all teams could understand and contribute to, the AI solution would not have a good foundation.”article, stated

Technical metadata can still fail business users if it omits business context.

Root cause: Schemas, lineages, and data types do not encode business definitions, ownership, usage context, or domain terminology by themselves.

Resolution: Unify technical and business metadata in a catalog exposed through APIs.

“Technical metadata includes table and column descriptions with data types, data lineage across transformations, synonyms and categorical indicators”article, stated

“Business metadata captures business definitions and rules for specific attributes, domain-specific terminology and ontologies”article, stated

Semantic-layer changes can break governed metrics if they are treated casually.

Root cause: Financial KPIs and SLAs depend on stable definitions, exemplars, and pipeline configuration.

Resolution: Use version control, CI/CD, staged deployments, gated approvals, and stakeholder sign-off for changes affecting KPIs or SLAs.

“An established change control process is essential for managing the dynamic nature of the data landscape”article, stated

“This approach requires stakeholder sign-off for changes that affect KPIs or SLAs while enabling safe, rapid deployment of improvements.”article, stated

Evaluation

The system uses exemplar question-to-SQL pairs as few-shot prompts, evaluation benchmarks, and regression tests, plus continuous evaluation mechanisms that track business metrics, user feedback, model degradation, and data reliability checks.

Metrics: over 50 exemplars · 3-5 business metrics recommended before launch · time-to-insight improvement from days to minutes · SQL query accuracy · analyst hours saved · user satisfaction · revenue or profit impact where measurable

3 supporting quotes

“Vanguard built a collection of over 50 exemplars that serve three purposes”article, stated

“The system maintains continuous regression suites and user feedback loops to evolve examples and semantics”article, stated

“Define 3–5 key metrics that matter to your business stakeholders and establish baseline measurements before launching your AI system.”article, stated

Outcomes

time-to-insight for complex financial queries dropped from days to minutes
business users gained independent access without SQL knowledge
metadata and semantic-layer implementation improved AI-generated SQL accuracy
routine analytical request workload decreased for the data team
the reusable framework began adoption across multiple Vanguard business units

The case reports enterprise-level adoption across multiple business units, with over 50 question-to-SQL exemplars and AI-ready data practices intended for regulated financial decision making.