What problem does it solve?
Top leagues now capture enormous amounts of live data: MLB says it is collecting over 15 million events and data points per game, and AWS reports that the Bundesliga's data foundation processes 200 million data points per match. Much of that data reaches fans only through the filter of a human broadcast commentary team, which exists for the biggest matches, in the main broadcast languages, and not for every match, market or language a league wants to reach.
A smaller league, a secondary market or a fan who does not speak the broadcast language gets the raw numbers on a stats page at best, not the narrative that makes those numbers meaningful in the moment. Producing that narrative live, in many languages, for many concurrent matches, is not practical with human commentators alone.
How does it work?
- Capture the live data. Tracking and event systems record the game at high frequency: ball, player and pose tracking, match events and historical statistics, arriving as a continuous feed while the match is live.
- Find the story in the data. A model scans the incoming feed for statistically interesting or narratively relevant moments, such as a record, a reversal or a repeated pattern, and turns each one into a short narrative point for a broadcaster or directly for fans.
- Generate the commentary. A generative model turns the selected data point or event into natural language in the target language, timed to the live action. Optionally, text to speech can voice the output for a spoken version.
- Publish alongside the live feed. The commentary appears next to the existing play by play in an app or web feed, or is delivered as an automated, multilingual live ticker running independently of the human broadcast.
- Monitor and correct. Editorial or product staff sample the generated commentary for accuracy and tone during and after live events, and can pull a specific feature if quality drops mid event.
- Audience
- Customer facing
- Autonomy
- Autonomous
- Adoption
- Early adopters
- Channels
- Mobile app
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has published a number for the KPIs of this use case yet.
Value drivers: Customer experience, Inclusion and access, Revenue growth.
Indicative value
A league with 380 matches a season and no live commentary in most fan languages
USD 11,400 to USD 1.5 million
Incremental streaming revenue from newly covered languages per season
How this is calculated
Formula: matchesPerSeason * fanLanguagesAdded * incrementalStreamsPerLanguage * revenuePerStream. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Matches per season matchesPerSeason, matches per season | 380 | 380 | The reference league, sized like a European top flight with 20 clubs playing home and away. |
| Additional languages automated commentary can cover beyond the broadcaster's own commentary team fanLanguagesAdded, languages | 3 | 8 | Editorial assumption. Replace with your own target market list. |
| Additional streams gained per match, per new language covered incrementalStreamsPerLanguage, streams per match | 500 | 5,000 | Editorial assumption, replace with your own audience research. |
| Average advertising or subscription revenue per incremental stream revenuePerStream, USD per stream | 0.02 | 0.1 | Editorial assumption for an ad supported or subscription stream. Replace with your own. |
What it leaves out: Gross incremental revenue only, before the cost of running the AI commentary service, any cannibalization of existing paid commentary products, and the risk that fans value human commentary for marquee matches in a way the automated version does not yet replace.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Bundesliga (DFL Deutsche Fußball Liga)
Germany · Media and entertainment · 2026
The Bundesliga uses an AWS data foundation that processes 200 million data points per match to power Data Story Finder, which surfaces narratives for broadcasters from match data, player statistics and tactical insights, delivered to commentators in milliseconds, and AI Live Ticker, which turns the same data into automated, multilingual commentary as a match happens.
AWS says AI Live Ticker is built on a fully serverless AWS architecture using Amazon Bedrock; Data Story Finder is described only as powered by the same AWS data foundation, not tied to Bedrock by name. The official Bundesliga app separately offers Captain, also built on Amazon Bedrock (and Amazon Nova), an AI companion that answers fan questions in natural language from official match data.
AWS's own marketing describes Captain, not AI Live Ticker, as agentic AI in production for 1B+ fans worldwide; that figure is AWS's framing of the Bundesliga's claimed global fan base, not a measured usage count for Captain, AI Live Ticker or any other product on this record.
No outcome disclosed.
Major League Baseball (MLB)
United States · Media and entertainment · 2026
Major League Baseball's Statcast system, built for scale, speed, security and flexibility using Google Cloud's Gemini Enterprise Agent Platform and BigQuery, analyzes ball, player and pose tracking data from every game and turns it into predictive models such as catch probability and steal success rate. Sean Curtis, MLB's Senior Vice President of Technology and Infrastructure, says MLB is collecting over 15 million events and data points per game.
Google Cloud reports that Statcast improves stats delivery by 300 milliseconds, enabling real time analysis of ball and strike calls. For the 2026 season, MLB launched Scout Insights, which Google Cloud describes as bringing AI generated color commentary and insight to the Gameday play by play feed in the MLB App and MLB.com.
No outcome disclosed.
Masters Tournament (Augusta National)
United States · Media and entertainment · 2023
For the 2023 Masters, IBM and the Masters Tournament added AI generated spoken commentary to the Masters app and Masters.com. IBM says the solution would produce golf narration for more than 20,000 video clips over the tournament, extending the MyGroup feature that lets patrons watch every shot from chosen players. IBM iX trained foundation models, including Watson Text to Speech, on the language of golf at the Masters, which automates the adding of spoken commentary to clips. The figure is a stated plan for the tournament, not a measured result.
No outcome disclosed.
The All England Lawn Tennis Club (Wimbledon)
United Kingdom · Media and entertainment · 2023
For the 2023 Championships, IBM and The All England Lawn Tennis Club introduced an AI Commentary feature that produces spoken commentary and toggleable captions for highlights videos on the Wimbledon App and wimbledon.com. IBM iX worked with the club to train foundation models from watsonx on the language of tennis. The club described it as a step towards offering commentary for matches outside the Show Courts, which already have live human commentary. The announcement gives no usage or accuracy figures.
No outcome disclosed.
United States Tennis Association (USTA)
United States · Media and entertainment · 2023
IBM and the USTA run AI Commentary on the US Open digital platforms: automated English language audio and subtitles for singles match summary highlight videos, first introduced in 2023. For 2024 IBM said it would use watsonx, including the Granite 13B model, for more frequent and contextual commentary, with highlight packages published minutes after a match ends. The same release adds AI generated Match Report summaries for all 254 main draw singles matches, reviewed by USTA editorial staff before publication. No accuracy or audience figures for the commentary are given.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A live, structured data feed of the event, such as ball and player tracking, match events and historical statistics
- A defined narrative style and vocabulary per sport and per language
- Historical commentary or narration to tune tone and style, where it exists
Systems to integrate
- Live sports data or tracking provider
- A low latency generation pipeline that can run per match, concurrently across many matches
- The publishing surface, such as an app feed, a live ticker or broadcast graphics
- Text to speech, for a spoken rather than written commentary
Complexity: High
Real time is the hard constraint: the system has to ingest a live data feed, decide what is worth saying, generate language, and sometimes speech, and publish it within seconds of the event, for every match running concurrently, in every target language. That is a different engineering problem from generating commentary after the fact.
- 1
Start with insight, not full commentary
Begin with short, data grounded insights, such as a record, a streak or a comparison, delivered next to the existing play by play, rather than trying to replace a full human commentary track on day one.
- 2
Fix the data feed before the language model
Live commentary is only as good as the tracking data behind it. Validate the accuracy and latency of the underlying data feed first.
- 3
Localize deliberately, market by market
Treat each language as its own launch with its own vocabulary, idiom and tone, rather than a direct translation of the main broadcast commentary.
- 4
Sample every live match for accuracy
Because commentary runs unattended in real time, review a sample of generated commentary from every match afterwards, not only when a fan complains.
- 5
Keep marquee events human led
Reserve human commentators for the matches or moments that carry the most commercial or emotional weight, and use automated commentary to extend coverage where a human commentary team does not reach.
Guardrails
- A fixed set of data sources and statistics the system may reference, so it cannot invent a statistic that is not in the feed
- Automatic suppression or delay if the live data feed itself looks wrong or disconnects, rather than continuing to narrate stale or invented data
- Human review of new languages and new sports before wide release
- Clear labeling that the commentary is AI generated
KPIs to instrument
- Fan reach per language, before and after automated commentary
- Accuracy of generated statements against the underlying data feed, sampled per match
- Latency from the live event to the published commentary
- Complaints or corrections per hour of generated commentary
Human in the loop
Editorial and product staff sample generated commentary from live matches for accuracy and tone, tune the narrative style before a new language or sport goes live, and can pull a specific automated feature mid event if the underlying data or the generated language goes wrong.
Common failure modes
- Confidently wrong commentary
- A generative model can state a plausible sounding but incorrect statistic if the underlying data feed lags or errors. Constrain the model to only state numbers it can trace to the live feed, and monitor for drift.
- Flat or repetitive narration at scale
- Running the same generation approach across many concurrent matches can produce generic, repetitive commentary that feels automated. Vary the narrative templates and sample output across matches, not only within one.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(2): a system that generates synthetic audio, text or video content, such as AI generated commentary, must ensure its output is marked in a machine readable format and detectable as artificially generated, unless a narrow exemption applies. That is a technical marking duty on the provider, not necessarily a visible on screen label for viewers; a visible disclosure that commentary is AI generated, as listed in controls below, is a design choice.
Article 50(4) can separately require a deployer to disclose that AI generated or manipulated text has been artificially generated, but only for text published with the purpose of informing the public on matters of public interest, and that duty does not apply where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publishing it.
Routine live match commentary will often sit outside Article 50(4) for both reasons: it is not usually framed as informing the public on a matter of public interest, and an editorial or product team typically reviews it. This is not an Annex III use unless the generated commentary itself were used to make a decision about a natural person, which is not the case in the deployments we found.
Rules that apply
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Generated audio, text or video content must be marked as artificially generated and detectable as such.
Controls to put in place
- Disclosure that commentary or insights are AI generated, in the app or feed itself
- A fixed list of data sources the system may cite, with no fabricated statistics
- Automatic fallback to silence or a generic message if the live data feed fails, rather than continuing to narrate
- Sampling review of generated commentary for accuracy after every event, not only on complaint
Frequently asked questions
- Can AI really commentate a live sporting event?
Today it is mostly used to add insight and multilingual narration alongside the main broadcast, rather than replace a lead commentator outright. MLB's Scout Insights adds AI generated color commentary and insight to its Gameday play by play feed, and AWS says the Bundesliga's AI Live Ticker turns match data into automated, multilingual commentary as events happen on the pitch. In 2023 IBM added AI generated spoken commentary to highlight videos at Wimbledon, to video clips at the Masters and to singles highlight videos at the US Open, where IBM said the 2024 packages would be published minutes after a match ends.
- How much data does this actually run on?
A lot. MLB says it is collecting over 15 million events and data points per game, and AWS reports that the Bundesliga's data foundation processes 200 million data points per match to power its live commentary and insight tools.
- Does AI generated commentary need to be labeled?
Under the EU AI Act, Article 50(2) requires the output itself to be marked in a machine readable format and detectable as artificially generated, unless a narrow exemption applies. That is a technical marking duty on the system, not automatically a visible label a fan sees on screen; showing one anyway is a sound design choice. Separate deployer obligations under Article 50(4) can apply, but only to text published with the purpose of informing the public on matters of public interest, and not to text that has had human review or editorial control by someone who holds editorial responsibility for it, so routine sports commentary will often fall outside that duty.
- What happens if the live data feed breaks mid match?
None of the deployments we found disclose this publicly. Treat a data feed failure as a guardrail case: the system should fall back to silence or a generic message rather than continue narrating stale or extrapolated data.
How to cite this page
Blits.ai AI Use Case Library, "AI generated live sports commentary and data storytelling", last verified 30 September 2026, https://www.blits.ai/ai-use-cases/live-sports-commentary-generation. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 12 November 2026: Quality check: set adoption to early adopters (five organizations run it in production), attributed the Bundesliga ticker description to AWS and added the IBM highlight commentary at Wimbledon, the Masters and the US Open to the first FAQ.
- 13 October 2026: Added deployments: The All England Lawn Tennis Club (Wimbledon), Masters Tournament (Augusta National), United States Tennis Association (USTA).
- 30 September 2026: Published after review by an automated review workflow (independent skeptic review).
- 30 September 2026: Editorial fix round. Reworded blitsAi.howToBuild: output guardrails are described as an LLM judged check against an admin authored policy (for example, blocking replies that state statistics), not as a check that compares a reply's numbers against the connected data source, since the guardrail never sees the data source; grounding a reply's numbers is now attributed to the custom function and agent instructions, tested with test suites and watched with monitors. Quoted Article 50(4) of the EU AI Act in full in faq and risk.euAiAct.basis: the duty applies only to text published to inform the public on matters of public interest, and does not apply to text that has had human review or editorial control by someone with editorial responsibility for it; noted that routine sports commentary will often fall outside it for both reasons, checked word for word against the archived regulation text (the EU's legal database blocks automated access, so we used the Wayback copy).
- 30 September 2026: Unpublished by an automated review workflow (independent skeptic review).
- 29 September 2026: First published