The most useful number in Formula 1's new AI case study is not the claimed 99% time reduction. It is the number of approvals the system kept.
F1 and Amazon Web Services say a new Data Accelerator cut source-onboarding work for F1's Customer 360 marketing platform from six to eight weeks to about 40 minutes of code generation, followed by hours of deployment and review. The striking design choice is that the agents do not bypass the engineering process. They manufacture evidence for it.
That makes the deployment a practical model for enterprise agents: automate the proposal before automating the decision.
The Workflow Is the Product
Customer 360 combines feeds from ticketing, streaming, sponsors, social media, and merchandise. According to the AWS/F1 technical account, each new source previously required manual schema mapping, ingestion code, quality checks, privacy classification, and governance policy. Twelve planned sources had accumulated into an 18-month backlog.
The new workflow starts with a business requirements document in Amazon S3. An agent running on Amazon Bedrock AgentCore generates configuration files, opens a GitHub pull request, and creates a linked Jira ticket. An engineer reviews that proposal.
Only after approval does a human trigger the next phase. The agent then creates three more pull requests: one for AWS Glue application and infrastructure code, one for DBT transformations, and one for governance policies including GDPR tagging. Engineers review those changes in the repositories they already use.
F1 and AWS report that agents now handle 95% of onboarding work, schema fixes take hours rather than days, and the backlog was cleared in weeks. Those are company-reported results, not an independent audit. They also preserve an important qualifier: the 40-minute figure covers code generation; deployment and review still take hours.
Copy the Four-Step Contract
The reusable pattern is Brief, Diff, Gate, Trace.
1. Brief
Give the agent a bounded input with named outputs. F1 uses a requirements document, not an open-ended instruction to “fix the data platform.” Constraint makes evaluation possible.
2. Diff
Require reviewable artifacts: configuration, tests, infrastructure definitions, transformations, and policy changes. A diff can be inspected, compared, rejected, and reverted. A vague autonomous action cannot.
3. Gate
Keep promotion outside the model. Every generated pull request receives engineer approval, access uses short-lived credentials, and the system runs inside private network boundaries. The agent proposes; deterministic controls and people decide.
This matters because the platform layer is not infallible. In a hands-on AgentCore review, InfoWorld found bugs and documentation problems in a sample deployment. It also observed a policy correctly deny an action while the model invented reasons for the denial. External policy worked; model self-explanation did not.
4. Trace
Bind each change to a ticket, pull request, test result, lineage graph, and rollback path. F1's observability layer connects pipeline health with business context and refreshes every 15 minutes. That turns automation from a black box into an accountable operating process.
Measure the Review Load
The next question is not whether an agent can generate 95% of the work. It is how much of that work reaches production unchanged.
Operators should track four measures:
- median human review time per generated change;
- correction rate before approval;
- rollback and incident rate after merge;
- cost and lead time per successfully onboarded source.
Without those numbers, code-generation speed can hide a new bottleneck in review. With them, teams can decide which change classes deserve tighter templates, automated checks, or eventually narrower approval authority.
The commercial pressure is real. Separately, TechRadar reported that TGR Haas's RaceMate AI fan companion saw up to 30% returning users, according to its technology partner. That is a different organization and deployment, but it shows why timely, governed fan data matters across the Formula 1 ecosystem.
The founder opportunity sits between agent demos and production systems: tools that turn ambiguous business requirements into policy-aware diffs, then prove what changed and why. Formula 1's lesson is simple. The safest place for an enterprise agent to start is not the driver's seat. It is the proposal lane.
Sources
- AWS and Formula 1, “From weeks to minutes: How Formula 1 uses agentic AI on AWS to accelerate data operations” (2026-08-03): https://aws.amazon.com/blogs/machine-learning/from-weeks-to-minutes-how-formula-1-uses-agentic-ai-on-aws-to-accelerate-data-operations/
- TechRadar Pro, “I want Racemate to be the second screen” (2026-07-28): https://www.techradar.com/pro/i-want-racemate-to-be-the-second-screen-infobip-is-helping-tgr-haas-to-fund-formula-1-car-development-through-an-entirely-new-kind-of-fan-engagement
- InfoWorld, “Running agents with Amazon Bedrock AgentCore” (2026-03-12): https://www.infoworld.com/article/4143387/running-agents-with-amazon-bedrock-agentcore.html