Scaling Content Production and Fan Engagement with Gen AI

bundesliga · Media & Entertainment · 2024 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (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: Sports content production and fan engagement with generative AI

Bundesliga needed to scale editorial content, image assignment, localization, and personalized fan engagement for a global football audience while keeping editors in control.

Constraints:

• The league serves a global fan base across many countries and languages.
• Editors need fast match reports, stories, images, and video localization during live sports cycles.
• Vector search over images was too latent and token-heavy for some match-report workflows.
• Fans face second-screen chaos during live matches and want statistical answers in natural language.
• Different fan questions need different model/workflow complexity for cost control.

Architecture

Bundesliga uses several GenAI workflows on AWS: match-report generation from OpenSearch/AWS Lambda match data with editor control and multimodal image assignment; story generation and image selection with Amazon Nova; video localization with translation review and Nova Pro; and Match Mate, a fan companion that uses dynamic text-to-SQL workflows, Athena/S3, OpenSearch video search, Amazon Nova Light/Pro, Anthropic/Sonnet routing, and proactive content delivery.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Content editor CMS"] --> B["Match data pipeline"]
  B --> C["Match report generator"]
  C --> D["Image assignment workflow"]
  C --> K["Human editor review"]
  E["Story generation workflow"] --> D
  F["Video localization workflow"] --> K
  G["Match Mate fan companion"] --> H["Dynamic workflow router"]
  H --> L["Amazon Nova models"]
  H --> M["Anthropic/Sonnet models"]
  G --> I["Athena SQL layer"]
  G --> J["Video OpenSearch retrieval"]
  K --> A

Components

Data flow

Editors trigger match-report generation; match data is transformed by Lambda into OpenSearch and used for prompts; LLMs generate text and select images from supplied image blocks; editors review output. For Match Mate, a fan question is classified by type and complexity, routed to Nova or Sonnet, transformed to SQL over Athena/S3, and paired with OpenSearch video retrieval when needed.

Why this architecture

Bundesliga uses specialized workflows rather than one static assistant: editorial workflows preserve human control, image blobs avoid unnecessary vector search, localization targets global reach, and dynamic routing keeps Match Mate cost-efficient.

Models used: Amazon Nova · Amazon Nova Light · Amazon Nova Pro · Anthropic · Sonnet 4 · LLMs for text-to-SQL

Scale

• Bundesliga has over 1 billion interested fans worldwide.
  “There are over 1 billion people over the world who are interested in the Bundesliga”transcript, stated
• The first GenAI solutions served more than 100,000 app fans.
  “served them to over 100,000 Bundesliga fans in our app.”transcript, stated
• Match report generation saved about 90% of editor time and enabled about 20 additional match-day articles.
  “save 90% of the time and use this time to generate, not generate but write other articles and other content pieces which amounts to about 20 additional content articles”transcript, stated
• Story creation from articles saved about 80% of editor time.
  “editors save about 80% of time in that endeavor.”transcript, stated
• Video localization processing time dropped by 75% and cost by three to five times.
  “reduces the processing time when localizing a video by 75%.”transcript, stated
• Match Mate scaled personalized content by five times per user and dynamic workflows reduced costs by 35%.
  “Match Mate enables Bundesliga to scale their personalized content by five times per user.”transcript, stated

Alternatives considered (roads not taken)

Vector search for match-report image selection

Rejected because: It added latency, token use, and image-count limits when the workflow already knew the relevant match images.

One static workflow for all fan questions

Rejected because: Question complexity varies, making a single path cost-inefficient.

International partners localize content manually

Rejected because: Manual partner localization limits distribution and global reach.

Patterns

model-routing (model-ops)

Match Mate routes simple and complex fan questions across Nova and Sonnet paths based on query type and complexity.

tool-use-orchestration (agent-architecture)

Match Mate orchestrates LLM text-to-SQL, Athena, S3 match data, and OpenSearch video retrieval.

human-in-the-loop-gate (governance)

Editors kick off, inspect, edit, and publish generated match reports and localized content.

cost-aware-routing (optimization)

Dynamic workflows and Nova price-performance reduce costs for fan-answer and image-assignment tasks.

Tradeoffs

Image assignment

Chose multimodal prompt with supplied image blocks over vector search over images.

Rationale: The known match image set avoided search latency, token waste, and image-limit problems.

Fan assistant workflow

Chose dynamic routing by question type and complexity over one static workflow.

Rationale: Different question complexities need different cost/performance paths.

Editorial automation

Chose editor-controlled generation over fully autonomous publication.

Rationale: Editors retain control over match reports and publication while AI removes repetitive production effort.

Gotchas (what bit them)

Traditional vector search was not ideal for match-report image selection.

Root cause: The workflow already knew the match image set, while vector search added latency and token costs.

Resolution: Supply image blocks directly to the LLM and ask for references.

Static fan-question workflows waste money.

Root cause: Fan questions differ by type and complexity.

Resolution: Classify query type and complexity, then route to different model paths.

Global fans prefer native-language content.

Root cause: Most Bundesliga content was produced in English or German despite a global audience.

Resolution: Build AI-powered video localization for international distribution.

Evaluation

Bundesliga measures editor time saved, story engagement, retention, localization time/cost, fan content scale, and dynamic-routing cost reduction.

Metrics: editor time saved · additional articles per match day · story app time spent · session lift · one-week retention · localization processing time · cost reduction · personalized content per user

Outcomes

• Match reports save about 90% of editor time.
• Story generation saves about 80% of editor time and Nova reduces a crucial image step cost by 70%.
• Video localization cuts processing time by 75% and costs by three to five times.
• Match Mate scales personalized content fivefold per user and reduces costs by 35%.

Bundesliga applies AI across a global fan base, production content workflows, localization, and interactive match-day engagement.