From weeks to minutes: How Formula 1® uses agentic AI on AWS to accelerate data operations

來源原標題

From weeks to minutes: How Formula 1® uses agentic AI on AWS to accelerate data operations

作者

Subhro Bose

原始標籤

Amazon Bedrock AgentCore · Amazon SageMaker Unified Studio · Artificial Intelligence · Customer Solutions · Intermediate (200) · Technical How-to

原始連結

https://aws.amazon.com/blogs/machine-learning/from-weeks-to-minutes-how-formula-1-uses-agentic-ai-on-aws-to-accelerate-data-operations/

摘要原文

Formula 1 ® (F1) engages an audience of over 800 million fans globally across digital platforms, F1 TV, social media, ticketing, and merchandise year-round. Races happen every two weeks. Fan engagement windows are measured in minutes and commercial decisions need to move at the speed of the grid. Behind the scenes, F1’s marketing technology (MarTech) platform, Customer 360, captures interactions across all of these touchpoints to power personalization, segmentation, and commercial strategy. However, the platform faced a significant operational challenge. According to Chris Roberts, Director of IT at Formula 1, “Our MarTech platform is the nervous system of F1’s fan engagement. But every new data source required 6 to 8 weeks of manual engineering. We had an 18-month backlog just to integrate 12 new sources.” The business was generating data faster than the engineering team could wire it up. As a result, Matt Kemp, F1 Head of Data Operations, set to improve efficiencies and data quality. “Manually ingesting data sources is time consuming, creates solution variances, and ultimately results in data integrity issues. I wanted a solution that was repeatable, robust and reliable. AWS worked backwards from our needs to implement an agentic solution that worked end to end, applying business logic at each step.” In early 2026, F1 and AWS worked together to build the Data Accelerator, a solution that uses agentic AI on Amazon Bedrock AgentCore to transform F1’s MarTech data platform from a manually maintained system into a self-managed, observable, and unified data estate. In this post, we show how the Data Accelerator reduced data source onboarding from up to 8 weeks to approximately 40 minutes of code generation plus hours of deployment. It also identified and fixed data source anomalies in production, tracked data platform operations and agent lineage in a single window, and opened a gateway for analysts, engineers, and scientists to collaborate. “For the first time, we have end-to-end visibility across the entire MarTech platform with data lineage and root cause analysis, not just dashboards full of alerts,” says Roberts. F1’s Customer 360 platform ingests data from ticketing partners, streaming integrations, sponsor activation feeds, social media, and merchandise systems. Operating a data estate of this breadth and velocity surfaced three areas of friction the team set out to solve. First, onboarding each new data source was a heavily manual effort: engineers wrote schema mappings, built ingestion pipelines, configured data quality checks, defined General Data Protection Regulation (GDPR) classifications, and set governance policies by hand. This process took 6 to 8 weeks per source. Second, the platform had to keep pace with constantly evolving upstream feeds. Providers frequently changed column names, added fields, or restructured and rescheduled payloads without notice. Those changes often surfaced at the worst possible moment, such as mid race-weekend or during a mission-critical campaign launch. Third, visibility was fragmented. Logs were scattered across services with no unified data lineage. When a stakeholder questioned a metric, engineers spent hours manually tracing the issue across Amazon Simple Storage Service (Amazon S3) paths, Amazon Redshift control tables, Airflow logs, and DBT outputs. The Data Accelerator addressed these challenges through five workstreams delivered simultaneously: A sixth workstream optimized the customer identity resolution algorithms that unify fan touchpoints across channels. The following sections describe each workstream in detail. The centerpiece of the Data Accelerator is a set of platform agents that take a Business Requirements Document (BRD) with limited information about the data source and produce a fully production-ready onboarding pipeline. This includes infrastructure code, data transformations, governance policies, and GDPR classification without a human writing a single line of boilerplate. The agents work in two phases: When a new data source needs onboarding, a team member uploads a BRD to an Amazon S3 bucket. The upload triggers an AWS Lambda function, which invokes Amazon Bedrock AgentCore Runtime, a capability of Amazon Bedrock AgentCore. The agent reads the BRD and generates a set of configuration files. It then accesses GitHub through a GitHub App to push these files as a pull request to the standardized Git repository, and accesses Jira through its REST API to create a ticket referencing the PR. All agent conversations and actions are traced in Amazon CloudWatch through built-in AgentCore observability. The assigned engineer reviews, adjusts if necessary, and approves. Phase 1 workflow: a BRD upload triggers the agent to generate config files and open a pull request Once the configuration files are approved, a human triggers the next stage. The agent takes the approved configuration and generates three separate Pull Requests: All three PRs link to a single Jira ticket for traceability. Engineers review each one across the Infrastructure, DBT, and Governance repositories and approve. Phase 2 workflow: the agent generates infrastructure, transformation, and governance pull requests What distinguishes this from a basic code generator is the integrated GDPR classification. The agent proactively analyzes every data column, determines whether it contains personal data, sensitive personal data, or pseudonymized data, and tags it with the appropriate GDPR category. These tags publish directly to the governance registry in SageMaker Unified Studio, giving the compliance team immediate visibility without manual review cycles. The system is not a tightly coupled agent graph. A single agent operates with modular skill definitions, each encapsulating a distinct capability: schema mapping and data type inference, data quality validation, governance enforcement, and sensitive data classification. At runtime, the agent evaluates incoming requirements and activates the relevant skills, composing them through a multi-pass reasoning process. Pass-0 handles token management through scrubbing, Pass-1 summarizes tool outputs, and Pass-2 rolls up an overall assessment, refining accuracy and completeness progressively rather than relying on a one-shot response. New capabilities ship as new skill modules without changing the core agent loop, keeping the architecture maintainable and composable as the platform grows. The result is onboarding time dropped from 6 to 8 weeks to approximately 40 minutes of code generation plus hours of deployment and review. AI agents now handle 95% of the work autonomously. Onboarding new data sources is one challenge, but keeping existing integrations healthy is another. Upstream providers frequently modify their data structures, from renaming a column to creating a new field. Previously, the F1 team discovered these changes when a pipeline failed, often during a live race weekend. The same agent architecture that handles onboarding now continuously monitors for upstream schema changes. When a provider modifies their data structure, the agent detects it through event-driven triggers using AWS Lambda and Amazon EventBridge. It assesses the downstream impact, identifying which pipelines are affected, and which consumers depend on the changed fields. It then generates the necessary code updates across all affected repositories and creates a Jira ticket with full context and linked PRs. Engineers receive a notification that explains what changed, describes the impact, and presents a proposed fix for review. End-to-end resolution now takes hours instead of days. Schema evolution agentic workflow Before the Data Accelerator, working with Customer 360 data required navigating multiple disconnected environments. Data engineers curated pipelines in one account. Data scientists who wanted to model fan behavior needed access to a separate account, and analysts operated in a third world entirely.