Formula 1 faced an 18-month backlog in integrating new data sources into their Customer 360 marketing technology platform, with each manual integration taking 6-8 weeks of engineering effort. In early 2026, F1 partnered with AWS to build the Data Accelerator, an agentic AI solution using Amazon Bedrock AgentCore that automates data source onboarding, schema evolution detection, and governance enforcement. The solution reduced onboarding time from weeks to approximately 40 minutes of code generation plus deployment hours (a ~99% reduction), with AI agents handling 95% of tasks autonomously while maintaining human oversight through pull request reviews. The system also provided unified data access through Amazon SageMaker Unified Studio and end-to-end observability with root cause analysis, ultimately clearing the 18-month backlog in weeks and improving data integrity across F1's fan engagement ecosystem.
Formula 1, engaging over 800 million fans globally through digital platforms, F1 TV, social media, ticketing, and merchandise, operates in an environment where races occur every two weeks and fan engagement windows are measured in minutes. Their Customer 360 marketing technology platform captures interactions across all touchpoints to power personalization, segmentation, and commercial strategy. However, the platform faced a critical operational bottleneck: each new data source required 6-8 weeks of manual engineering work for schema mapping, pipeline construction, data quality checks, GDPR classification, and governance policy configuration. This resulted in an 18-month backlog just to integrate 12 new sources, with the business generating data faster than engineering could integrate it. The manual nature of this work also created solution variances and data integrity issues.
In early 2026, F1 partnered with AWS to build the Data Accelerator, a comprehensive agentic AI solution that fundamentally transformed their MarTech data operations from a manually maintained system into a self-managed, observable, and unified data estate. The solution demonstrates sophisticated production use of LLMs through Amazon Bedrock AgentCore, representing a mature example of agentic AI deployed in mission-critical data infrastructure.
The Data Accelerator is built on Amazon Bedrock AgentCore, which hosts agents in runtime containers with long-term memory capabilities that retain context across invocations. The foundation model used is Claude Sonnet 4.6, accessed through Amazon Bedrock. Development leveraged Kiro for structured spec-driven development, providing a disciplined framework for building the agentic system.
The broader event-driven architecture integrates AWS Lambda for compute, Amazon EventBridge for event routing, Amazon Managed Workflows for Apache Airflow (MWAA) for workflow orchestration, and Amazon S3 as the raw data layer. All AI model access is governed through F1’s AI Gateway, which provides unified access control, cost management, and audit logging. This gateway approach represents a production-grade pattern for managing LLM access at scale, ensuring that every model invocation is tracked, attributed, and compliant with organizational policies.
The entire system operates within private subnets in Amazon VPC with no direct internet access, and all credentials are encrypted at rest in AWS Systems Manager Parameter Store. This network isolation and credential management approach demonstrates enterprise-grade security posture for production LLM deployments.
A key architectural decision distinguishes this solution from simpler code generation tools: rather than implementing a tightly coupled agent graph, the system uses a single agent operating with modular skill definitions. Each skill encapsulates a distinct capability including schema mapping and data type inference, data quality validation, governance enforcement, and sensitive data classification. At runtime, the agent evaluates incoming requirements and activates relevant skills, composing them through a multi-pass reasoning process.
The multi-pass reasoning approach is particularly sophisticated. Pass-0 handles token management through scrubbing, Pass-1 summarizes tool outputs, and Pass-2 rolls up an overall assessment. This progressive refinement improves accuracy and completeness rather than relying on a one-shot LLM response, representing a mature pattern for production agentic systems where reliability matters more than raw speed. New capabilities can be shipped as new skill modules without changing the core agent loop, maintaining architectural composability and maintainability as the platform evolves.
The agentic data source onboarding process operates in two distinct phases. When a new data source needs onboarding, a team member uploads a Business Requirements Document (BRD) to an Amazon S3 bucket. The upload triggers an AWS Lambda function, which invokes Amazon Bedrock AgentCore Runtime. The agent reads the BRD and generates a set of configuration files, then accesses GitHub through a GitHub App to push these files as a pull request to the standardized Git repository. The agent also accesses Jira through its REST API to create a ticket referencing the PR, establishing full traceability from the initial request through to deployment.
All agent conversations and actions are traced in Amazon CloudWatch through built-in AgentCore observability, providing comprehensive visibility into agent decision-making and tool usage. The assigned engineer reviews the generated configuration, adjusts if necessary, and approves. This human-in-the-loop pattern represents the “human at the helm” principle that gives F1 confidence to deploy agentic AI in production data pipelines.
Once configuration files are approved, a human triggers the next stage. The agent takes the approved configuration and generates three separate pull requests: AWS Glue application and infrastructure code, DBT transformation framework, and governance policies including GDPR tagging. All three PRs link to a single Jira ticket for traceability. Engineers review each one across the Infrastructure, DBT, and Governance repositories before approval.
This separation of concerns across multiple repositories reflects production data engineering best practices, where infrastructure, transformation logic, and governance are managed independently but orchestrated together. The agent must understand the relationships between these concerns and generate coherent, compatible code across all three domains.
What elevates this beyond basic code generation is the integrated GDPR classification capability. 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 Amazon SageMaker Unified Studio, giving the compliance team immediate visibility without manual review cycles.
This automated governance enforcement represents a critical production pattern for LLM-based systems handling regulated data. The LLM’s natural language understanding capabilities enable it to reason about data semantics and privacy implications in ways that traditional rule-based systems struggle with, while the declarative configuration approach ensures that governance decisions are codified, auditable, and consistently enforced.
Beyond initial onboarding, the same agent architecture 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 downstream impact, identifying which pipelines are affected and which consumers depend on changed fields. The agent then generates necessary code updates across all affected repositories and creates a Jira ticket with full context and linked PRs.
Engineers receive a notification explaining what changed, describing the impact, and presenting a proposed fix for review. This reduces schema evolution resolution from days to hours. The proactive detection and remediation capability demonstrates how agentic AI can provide continuous operational support beyond one-time code generation tasks, functioning as an always-on platform reliability partner.
The solution implements Amazon SageMaker Unified Studio as the foundation for a data mesh framework where a central governance account brokers data discovery and access across multiple producer teams. The key enabler is that governance is codified as declarative configuration, not manual console operations. A single data source definition simultaneously publishes data to the catalog and provisions the access control needed for consumers to subscribe.
This declarative governance approach is what makes safe, autonomous agent operation possible. Agents can onboard new data products end-to-end, from storage to catalog to governed access, because the framework enforces security constraints by construction. No human needs to review IAM policies or AWS Lake Formation grants because the platform guarantees correctness structurally. This represents a sophisticated production pattern where LLM-based automation is only safe when the underlying platform provides strong guardrails through declarative policy.
Data engineers curate and govern datasets in one environment, and data scientists find those same datasets in the same environment: governed, documented, and ready to model. A data scientist building a fan segmentation model doesn’t need to know where data lives, who owns the pipeline, or which S3 prefix to use. They open Unified Studio, find curated Customer 360 datasets, and start modeling with shared notebooks, consistent tooling, and governed access.
The observability layer presents full data lineage from S3 raw ingestion through processed layers into Amazon Redshift DBT stages as a single interactive graph, color-coded for health. Users can click on any node to drill down to individual sources and tables, each showing pass/fail status, last run time, and duration. When a pipeline fails, the lineage visualization shows exactly where the break occurred and which downstream data is affected.
Root Cause Analysis (RCA) is an agentic tool within F1’s platform that reads system logs and identifies failure points across the data estate. What makes this particularly sophisticated is that RCA is augmented by passing through business context and system topology, codified as JSON. A missing file in S3 might be the technical error, but with the context graph, RCA can explain that the upstream provider rescheduled their delivery window, which is why the file wasn’t there when the pipeline ran. This represents the difference between knowing what failed and understanding why.
The observability dashboard auto-refreshes every 15 minutes and provides, for the first time, full lineage, causal root cause analysis, and business context definitions in one place. This unified observability powered by agentic AI represents a significant operational advancement over fragmented logs across services.
A sixth capability layer involves automated identification of failures in the observability dashboard and agentic operation to fix them if they can be addressed with code changes. This closes the loop from detection to diagnosis to remediation, creating a self-healing data platform capability. The agent doesn’t just generate code during initial onboarding; it actively participates in ongoing platform operations, identifying issues and proposing fixes.
This operational automation represents a mature LLMOps pattern where agents move beyond assistive tooling to become active participants in production system reliability. The human-in-the-loop approval process ensures that no automated change reaches production without review, maintaining the “human at the helm” principle even in this autonomous operational context.
The system operates on the principle that AI proposes and humans review. Every generated pull request goes through engineer approval using the same review process the team already uses, ensuring immediate adoption without process disruption. The security posture includes least privilege access with fine-grained permissions and short-lived tokens with one-hour expiry, access limited to specific repositories and resources, full audit trail with every action logged and attributed for compliance, comprehensive automated testing where agents generate tests for their own changes, and rollback capabilities for issues surfaced post-merge.
This multi-layered security approach demonstrates production-grade LLMOps practices. The short-lived tokens with one-hour expiry, in particular, represent a sophisticated approach to credential management for agentic systems that need programmatic access to critical resources. The fact that agents generate tests for their own changes shows how LLM capabilities can be leveraged not just for primary code generation but for the entire quality assurance process.
A separate workstream optimized the algorithms that resolve customer identity across F1’s fan touchpoints. A single fan might interact through the app, buy tickets on the website, watch on F1 TV, and engage on social media. Unifying those interactions into a single identity without false merges or missed matches enables effective personalization within the Fan Personalization Platform.
Rather than re-architecting the pipeline or replacing components, the team focused on optimizing the existing resolution algorithm’s computational performance. By profiling execution bottlenecks and tuning the matching logic, the engagement reduced processing time by 50% while keeping the entire resolution pipeline and downstream integrations fully intact. With faster resolution, F1 can onboard any new data source and gather new customer data in half the existing time, enabling more timely and relevant personalization across every marketing channel.
The Data Accelerator delivered measurable impact across F1’s MarTech operations. Data source onboarding was reduced from 6-8 weeks to approximately 40 minutes of code generation plus hours of deployment and review, representing approximately 99% time-to-value reduction. AI agents now handle 95% of onboarding tasks without human intervention. Schema evolution resolution moved from days to hours. The 18-month integration backlog was cleared in weeks. Data engineers who previously spent time writing boilerplate ingestion code and chasing schema breaks now focus on strategic initiatives.
Remarkably, a single developer took the agentic solution from proof of concept to production release in 4 months, demonstrating the velocity achievable with modern agentic AI frameworks when properly architected. The reliability, consistency, and data integrity of the MarTech platform improved while operational overhead decreased.
While the case study presents impressive results, readers should consider several factors when evaluating this approach. The solution represents a significant partnership between F1 and AWS, with acknowledgments listing numerous AWS contributors. This suggests substantial professional services engagement and AWS expertise was required to achieve these outcomes. Organizations without similar access to vendor expertise may face different implementation challenges.
The “approximately 40 minutes of code generation plus hours of deployment” framing requires careful interpretation. While dramatically faster than 6-8 weeks, the “plus hours of deployment” component and the human review cycles mean this is not fully autonomous infrastructure provisioning. The human-in-the-loop approval process is presented as a security feature, which it certainly is, but it also indicates that the agents cannot yet be fully trusted to make production changes autonomously.
The modular skill architecture and multi-pass reasoning approach are sophisticated and likely represent significant engineering investment beyond simply connecting an LLM to APIs. The declarative governance framework that makes safe agent operation possible required building a platform layer that enforces correctness structurally. Organizations considering similar approaches should recognize that the LLM is only one component in a broader platform architecture.
The case study does not discuss model costs, token usage patterns, or economic tradeoffs in detail. At production scale with continuous schema monitoring and automated RCA, LLM invocation costs could be substantial. The use of Claude Sonnet 4.6 suggests prioritizing capability over cost optimization, which may be appropriate for this use case but represents a specific design choice.
The security posture with short-lived tokens, least privilege access, full audit trails, and network isolation represents production-grade practices that should be considered essential for any similar deployment. The fact that F1 built an AI Gateway for unified access control and cost management suggests these concerns are real and require dedicated infrastructure.
The approach is deliberately designed to be replicable. The agents are domain-agnostic in that they know how to onboard, classify, and monitor, but the domain knowledge is interchangeable. Any organization dealing with multi-source data onboarding, schema volatility, and governance requirements could apply similar architecture patterns to their own environment.
The key architectural principles that make this replicable include using modular skill definitions rather than monolithic agent implementations, implementing declarative governance that agents can safely operate within, maintaining human-in-the-loop approval for all production changes, using event-driven triggers for continuous monitoring rather than just one-time automation, and building comprehensive observability into agent operations from the beginning.
The success of this implementation demonstrates that agentic AI has reached production readiness for complex data engineering workflows when properly architected with appropriate safeguards, governance frameworks, and human oversight. The “human at the helm” principle appears to be the critical enabler that gave F1 confidence to deploy this in mission-critical marketing infrastructure, suggesting that fully autonomous agent operation may not yet be the goal for most enterprise LLMOps deployments.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
Novartis partnered with AWS Professional Services and Accenture to modernize their drug development infrastructure and integrate AI across clinical trials with the ambitious goal of reducing trial development cycles by at least six months. The initiative involved building a next-generation GXP-compliant data platform on AWS that consolidates fragmented data from multiple domains, implements data mesh architecture with self-service capabilities, and enables AI use cases including protocol generation and an intelligent decision system (digital twin). Early results from the patient safety domain showed 72% query speed improvements, 60% storage cost reduction, and 160+ hours of manual work eliminated. The protocol generation use case achieved 83-87% acceleration in producing compliant protocols, demonstrating significant progress toward their goal of bringing life-saving medicines to patients faster.
Prudential Financial, in partnership with AWS GenAI Innovation Center, built a scalable multi-agent platform to support 100,000+ financial advisors across insurance and financial services. The system addresses fragmented workflows where advisors previously had to navigate dozens of disconnected IT systems for client engagement, underwriting, product information, and servicing. The solution features an orchestration agent that routes requests to specialized sub-agents (quick quote, forms, product, illustration, book of business) while maintaining context and enforcing governance. The platform-based microservices architecture reduced time-to-value from 6-8 weeks to 3-4 weeks for new agent deployments, enabled cross-business reusability, and provided standardized frameworks for authentication, LLM gateway access, knowledge management, and observability while handling the complexity of scaling multi-agent systems in a regulated financial services environment.
