Use Case Intelligence RSS Search use cases Search
Databricks Software / SaaS United States Company newsroom

Capital One Software

IT Operations Automation

Capital One Software found that a static GenAI cost dashboard could show spend and governance KPIs, but could not answer the pointed follow-up questions teams had about model usage, tagging gaps, error costs, chargeback readiness, and savings opportunities.

Read the original story Explore all use cases
20 Unity Catalog SQL functions
20 functions across 5 domains
20 tools
6 KPI counters, twelve charts, four detail tables

Reported impact

Business outcome

Capital One Software replaced a fixed dashboard with a conversational GenAI cost analyst that can answer ad hoc FinOps questions, surface governance gaps, rank remediation opportunities such as untagged or dormant endpoints, and provide a governed path from prototype to monitored production deployment.

  • Cost Reduction
  • Productivity Improvement
  • Risk Reduction

Solution

Capital One Software implemented a GenAI FinOps agent on Databricks by turning Databricks System Tables into a reusable knowledge layer. Billing, list-price, endpoint-usage, and served-entity data from system.billing.usage, system.billing.list_prices, system.serving.endpoint_usage, and system.serving.served_entities are exposed through 20 Unity Catalog SQL functions registered in a dedicated schema, documented with COMMENT metadata, and validated by tiered smoke tests. A Supervisor Agent in Databricks AI Playground uses those Unity Catalog tools to answer free-form questions about spend, governance gaps, attribution, errors, and model optimization, including token-share cost allocation across users and models. For production, the same design is exported to Databricks Apps, where agent.py uses the OpenAI Agents SDK, Databricks MCP servers or DatabricksFunctionsClient connect back to the registered functions, and MLflow Agent Server plus MLflow tracing provide deployment, observability, and monitored operation.

Data flow

Billing, price, endpoint-usage, and served-entity telemetry flow from Databricks System Tables and AI Gateway-backed serving data into 20 Unity Catalog SQL functions in a dedicated schema. The Supervisor Agent in AI Playground calls those functions as tools, combines cost, governance, attribution, and error analyses, and returns natural-language answers. In production, Databricks Apps uses MCP-based function calls while MLflow records traces for each interaction.

Solution architecture

11 components · 4 layers
  1. Storage
    • Databricks System Tables Provide the billing, list-price, endpoint-usage, and served-entity telemetry that the agent queries as its knowledge base.
  2. Serving
    • AI Playground Serves as the built-in interface where the team prototypes and chats with the tool-calling cost supervisor.
    • AI Gateway Supplies serving telemetry and fallback or traffic-splitting behavior that feeds the attribution logic through serving tables.
    • Databricks Apps Packages the prototype into a deployable production application exported from AI Playground.
    • MLflow Agent Server Exposes the deployed agent through an async /invocations endpoint with tracing and streaming response support.
  3. Orchestration
    • Supervisor Agent Reasons across the registered functions to answer spend, governance, attribution, and optimization questions.
    • OpenAI Agents SDK Provides the agent framework used in agent.py before the app is wrapped for Databricks-compatible serving.
    • Databricks MCP servers Handle tool discovery, execution, and security between the production agent and Unity Catalog functions.
    • DatabricksFunctionsClient Wraps the registered functions for the production connection pattern used in the experiment.
  4. Governance
    • Unity Catalog Hosts the registered SQL functions, their comments, and the permissions the agent relies on for governed tool access.
    • MLflow Captures traces, experiments, and observability data for each deployed interaction.

Architecture clues

  • Each function takes a lookback_days parameter and returns a table.
  • Functions are registered with CREATE OR REPLACE FUNCTION and validated with tiered smoke tests.
  • MLflow Agent Server provides an async FastAPI server with an /invocations endpoint.
  • Production export records every interaction as a trace in an MLflow experiment described as the agent's flight recorder.
  • The agent supervisor configuration is limited to 20 tools.
  • Token-share allocation uses a three-CTE pattern: endpoint_cost, user_tokens, and totals.

Evidence from the source

"Allocated_cost = endpoint_daily_cost x (user_model_tokens / total_endpoint_tokens)"
"Every interaction with the deployed agent—every tool call, every response—gets recorded as a trace in this experiment."
"How we turned Databricks System Tables into a knowledge base for an AI agent that answers any GenAI cost questions on demand."
"It creates a schema, registers all 20 Unity Catalog SQL functions with CREATE OR REPLACE FUNCTION and runs tiered smoke tests to validate each function against your actual data."
"No dashboard is going to be able to anticipate every question, but an AI agent can answer any of them, if it has the right knowledge layer to refer to."
"The entire agent runs in AI Playground, Databricks’ built-in interface for prototyping tool-calling agents."
Transparency Evidence & methodology 90/100 reference value · 97% confidence · 2 sources

Review how this reference was scored and trace each normalized claim back to its documented source.

01

Scoring methodology

How scores are calculated

Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.

Customer reference value90/100
  • Problem clarity20/20
    Measures whether the customer's operational problem is specific enough to compare.
  • Implementation depth25/25
    Rewards named components, data flow, integrations, and architecture detail.
  • Outcome specificity20/20
    Measures whether the reported benefit is concrete rather than a generic claim.
  • Quantified impact15/25
    1 metric describe measurable impact.
  • Deployment scale10/10
    Rewards evidence of a live or scaled deployment and meaningful integration scope.
Evidence confidence97%
  • Source authority25/25
    The deployment is described by the adopting organization or a government source.
  • Claim traceability23/25
    11 of 12 key fields link to a named source.
  • Direct evidence20/20
    6 direct source quotes retained.
  • Date provenance10/10
    The publication date comes from the source page or publisher.
  • Source transparency9/10
    2 source references document publisher, URL, authority, retrieval, and supported fields.
  • Implementation specificity10/10
    Rewards source-backed components, data flow, quantified impact, and a specific outcome.
02

Traceability

Sources & provenance

Field coverage shows which facts each source supports. Dates are shown only when documented by the source.

  1. Capital One Software

    Capital One Software customer story

    First-party
    https://www.capitalone.com/software/blog/databricks-genai-cost-supervisor-agent/
    Source type
    Customer-published story
    Published
    Mar 19, 2026
    Supports
    • Customer
    • Business problem
    • Solution
    • Products
    • Solution components
    • Data flow
    • Integrations
    • technical area
    • use case category
    • Business outcome
    • outcome category
    • Metrics
    • Architecture
    • evidence quotes
    • Publication date
  2. Capital One

    Corporate Offices | Capital One

    First-party
    https://www.capitalone.com/about/corporate-information/corporate-offices/

    The page says, "Capital One is headquartered in McLean, Virginia" and lists "United States".

    Source type
    Official company website
    Supports
    • Region

Date evidence

  • Published on source page Mar 19, 2026 · Capital One Software · First-party
    Open date source
Not separately sourced Industry, Company size. Treat these normalized attributes as lower-confidence until a dedicated source is added.

Related

Similar use cases

See all like this
Databricks Jan 21, 2026 Mirakl Software / SaaS France Mirakl needed a better way to support generative AI supplier catalog onboarding for ecommerce marketplaces because strict governance rules and siloed data teams slowed marketplace analytics and made supplier catalog… 30x faster iteration and delivery across data and ai teams Document ProcessingIT Operations Automation Agent Bricks · Databricks Data + AI Platform · Lakeflow Jobs Google Cloud Archived · Oct 16, 2026 Alta Software / SaaS United States Alta needed a transactable Google Cloud Marketplace path for its AI and technology spend management platform without pulling its product team off core forecasting work, while also helping finance and IT leaders… Saved 125 engineering and 60 project management hours IT Operations AutomationReal-time Analytics Cloud Pub/Sub · Cloud Run · Cloud SQL Alibaba Cloud Oct 12, 2026 ECLICKTECH (易点天下) Software / SaaS China ECLICKTECH needed to industrialize AI-driven advertising operations, but media-delivery workflows were still long, manual, and heavily dependent on senior operators for anomaly diagnosis and creative adaptation. As more… AI covers about 50% of core execution work IT Operations AutomationMarketing Personalization AdClaw · AgentCore · AgentLoop