🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - AI Agents Slash PGA TOUR's Content Cost While Boosting Coverage/Quality -SPF204
In this video, David Provan from PGA Tour and Murali Baktha from AWS explain how they implemented agentic AI to generate content at scale for PGA Tour's digital platforms. They detail the architecture using AWS Bedrock and Agent Core runtime to produce 140-180 articles weekly, with plans to reach 800 articles per week. The system generates betting profiles, tournament recaps, and player analyses for 156 players per tournament, achieving 95% cost reduction at 25 cents per article while delivering content within 5-10 minutes post-tournament. The solution uses multiple AI agents including research, writer, editor, and validation agents working with both structured data from APIs and unstructured data from media guides. They emphasize production-first thinking, automated validation, and maintaining brand guidelines while generating billions of page views annually. The presentation also covers their broader AI strategy, focusing on using AI where it provides genuine value rather than forcing AI into every project.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Introduction: The PGA Tour's Digital Challenge and Data-First Strategy
Good morning everyone. Thank you so much for joining us this morning, a full three hours into the event, so I assume the doughnut sugar high is kicking in any second now. My name is David Provan. I'm Vice-President of Digital Architecture with the PGA Tour. I'm here with Murali Baktha, a Solutions Architect from AWS, and we're here today to talk to you about the work we have done at the PGA Tour on using agentic AI to write content for distribution on our digital platforms. I'm going to stand up because it's awkward to sit down. I just feel weird about it.
Let's start by describing what the PGA Tour is, because some people don't know and that's okay. The PGA Tour is responsible for operating 40 golf tournaments throughout the globe at this point for our players. Unlike other sports leagues, we're actually player owned. Many other sports leagues are owner driven, but our players actually own the PGA Tour. Our job is to deliver golf tournaments for the players to have playing opportunities and improve their game and earn prize money essentially. We do that pretty much every week of the year, and then we get like three weekends off. Other than that, there's golf every weekend.
We also operate four different tours. You may have all heard of the PGA Tour, we have the Corn Ferry Tour, which is our qualifying event, PGA Tour Champions, and PGA Tour Americas. These are all opportunities for our players to be seen, to be engaged with, and to play the sport they want to play to progress up to the PGA Tour. We're a bit unlike other sports. Other sports you play inside white lines, or you play in a stadium or a fixed area. We don't do that. We play in these amazing venues which are 150 acres plus. Everything you see is playable area. There is not an inside this white line, stay there mentality. We've had players purposely hit balls on top of clubhouses to get relief. We will have John Deere tractors in the middle of the golf course when it's a John Deere Classic. So we have natural objects that players are working around to play golf, and all of that is part of our playable area.
To data collect in that is a real challenge for us. We have 156 players, 31,000 shots in a standard golf tournament in stroke play. We then decided that wasn't hard enough, so why don't we do different scoring formats. Some weeks you'll watch team stroke play, match play, Stapleford scoring, so we have a whole variety of variables that go into what we do. Our biggest challenge inside all that is how to engage you guys on what we're doing. Golf is hard. Golf is sometimes particularly hard to watch live, but sometimes it's hard to know what the heck is going on and why should I care. So we're building agentic systems to help us solve those things, and we're going to cover those today.
If you look at our digital strategy, at the core of what we do, our website is the base layer. It's got everything you could possibly want. To the nth level, if you want to golf nerd out, it's probably on PGA Tour.com somewhere, and you'll find it and live your best life. Our mobile apps are highly engaged platforms, and we see a seven times consumption compared to web, but it's highly focused like short-term use in, leaderboard out. In leaderboard, watch videos out, but users will come back seven or eight times and use the platform a lot. More recently we've played with Vision Pro and other platforms, and we have this kind of content strategy where we will build out from that.
We've actually been successful working with other golf organizations more recently in bringing some of this PGA Tour technology, whether it's Shot Link or digital products like Torcast and delivering them for USGA and PGA of America to create that experience that fans have become dependent on. So like I said, we're data first. The biggest problem we have is gathering data off a golf course and making it relevant to our players, to our sponsors, and to our fans. That's the challenge we face every week, and I do mean every week on the PGA Tour.
Capturing Petabytes of Data: Shot Link Technology and the PGA Tour Data Lake
We are gathering petabytes of data. If you come to a PGA Tour event, which you should all come, please, you're welcome. You'll see these cameras, which are like the world's best security camera system. We have 144K cameras dotted around the golf course. They are scanning each one at five frames per hole, and they're analyzing the ball data as it comes in. But we also have two radars on every single hole, one on the tee, one on the green. They are military grade radar used for tracking tiny white objects, and in our case, our tiny white objects are golf balls. We've also got on there the cameras all plug in together, and we're starting to look at player movement, where the players are, and we thought we'd built this really great golf ball tracking system and our security team were like, that's also a really great closed circuit camera system for out of hours, can you leave them turned on please?
Aside from all of that other data, we have walking scorers, capturing when the ball is hit, ball in motion, and if you watch a Shot Link event underway, the determination of the Shot Link team to capture every single shot is a militant determination to be perfect. It's incredible to see those guys determined that we will not miss a single shot.
We deliver those to gaming providers, we look to digital platforms, but we capture a very large and significant amount of data on site. We also build solar infrastructure. If you come to a PGA Tour event and you see the cameras, at the bottom of the camera you're going to see a black skirt. If you get really close to it, which you can because the playable area is also the area you get to walk in, you'll see solar panels have been stitched into the skirt. So we can run our cameras through solar panels, we get two or three days of light, which we do on golf courses typically. We can run that, and we're not shipping power and stuff, so at our core, we're really an operational team that runs and deploys golf tournaments.
Like I said, 150 acres, 140 4K cameras throughout each course, 156 participants each week, and on average 14 balls in the air simultaneously, and we've covered this stuff. So where do we store that? The first project we undertook with AWS was this data lake. Where do we put the data, and we built this massive collection of scoring data, ball data, content, all the stuff we could possibly need, and it's highly structured. We know every single shot back to, I think, 2020 is our shot data collection, might be before then. We have scoring and leaderboard data back to 1864, if you'd like it. So we have tons and tons of data.
Our real challenge is how do we storytell with that data? How do we pick what's relevant at the right point, at the right time, and the right mechanism to make you engage with our platform. On top of that, we've also built a media lake. On a given standard PGA Tour event, we're running seven live video streams, four with ESPN, some with the Golf Channel, some with NBC, some with CBS. Those are all stored in the media lake. We have thousands of hours of golf coverage, which we've essentially built a shopping cart. You can come and say, I want all Scottie Scheffler putts that are less than ten feet at this golf course, and we'll create a playlist that we can give to our broadcasters and say there's your supporting reel you can use in broadcast.
Product-First Approach to AI: Balancing Fans, Business, and Brand
So this data lake and media lake is super important and kind of the core foundation of what we're building at the tour. Now we'll talk about the AI side and our challenges for the tour in terms of what we deliver. We started looking at AI about two to three years ago, and I think much like all of us went, what do we do with it? Sounds cool, not sure what to do with this, not sure how to drive value out of it, not sure how we use it properly. So actually we have a really strong product team in our digital team. We took a product-first approach in what we did.
So we build any feature for PGA Tour digital platforms. What's the value for the fan? What are the values to the stakeholders and what are the values to the brand? Those three things are at the top of what our product team does week in, week out. We just apply that to AI. What can we do that will help our fans consume content, provide more coverage, and be in the PGA Tour voice and respect that brand? So like I said, fans, business, and brand. Sometimes we do things that match all three of these, sometimes we do things that match one of them, sometimes we do something because it's cool. It's not fan engagement, we're a sports league, you want to look good, and sometimes we do things that no one sees, but it cuts my AWS bill by twenty percent. It's important to the business, right?
As we build those features, we then have this production-first thinking in what we're doing. I have worked in technical sports for about sixteen years. We live in the edge case. We are really good at edge cases. What happens if someone holes in one and damages the hole and we've got to repair it in the middle of a round? What happens if Jordan Spieth puts the golf ball on top of a clubhouse to get free relief? All these things seem ridiculous, but that's what we call Thursday. We have these thirty-one thousand shots. They are not uniform, they're not consistent. Golf is hard for those of you who play it. So we live in things that can go wrong. So we have to bring that kind of edge case thinking to the forefront of what we do.
The first thing we talk about is how do we score and validate. We work in sport, we can't be second. If we're second, we're last. We have to be first. So how do we automatically improve things as quickly as possible using scoring and validation? Our features essentially will go through an analysis and work out if you've passed the validation techniques. Some of it's simple rejects, some of it's LLM as a judge, some of it's take the original data and check it's in the output. There are different techniques we use there. Sometimes we'll throw it to a human. We're building a new feature, as Murali will come onto. We'll have a few running drafts, we'll verify it and we're happy with the quality. We then flip the auto button on.
And then more importantly we plan for that production at the beginning. All the features that Murali will take you through here, we have not changed our support footprint. We've kept the same support team. We've extended AWS CloudWatch dashboards and integrations operationally to allow us to monitor those systems and prove that they're operating at the level we expect and in fact demand. So we don't want to be in the business of watching a golf tournament. I've always said my favorite idea of success is if we sit there and twiddle our thumbs for four days of a golf tournament, that's a giant check mark. We succeeded. The best analogy I have is the golf tournament may look like a swan above the river, but the legs are going like crazy, and hopefully none of you see that. If none of you see it, we've done a good job.
From Chatbots to Shot Commentary: Launching AI-Generated Content for Fan Engagement
This journey we took into AI was a really interesting one, because we pivoted along the way. At the start of 2023, we worked with PGA Tour. Everyone wanted a chatbot. Full confession, David doesn't like chatbots. I think chatbots are a great way to make your brand be exposed. It's just that you have to really confine what they can do.
We played with chatbots, then we learned from that and started to pivot. We asked what are the things that can make a difference. Through 2024 we prototyped and worked on this idea of shot commentary. For those of you that follow sport, if you watch NFL on a Sunday or Bundesliga or the Premier League, the commentator is quite important. The color commentator is really the icing on the cake that sets the context for you. It's not just the play by play, but why does that play matter.
At Players 2025, we launched our shot commentary system. The shot commentary system works in two aspects. It generates a fact: Scotty Scheffler has hit the ball 164 yards, he is 5 feet from the pin. It then adds context: He has a 96% chance of making this putt. If he birdies it, he's going to move to position one on the leaderboard and take the overall lead of the FedEx Cup. That is actually useful. We were really strong that we didn't want to do narration, we wanted to do commentary. We held ourselves accountable to a really high standard where it can be repetitive, it's got to be varied, it's got to be engaging, it's got to matter, which also means sometimes we don't do the context.
Sometimes it's just his first drive on the first hole and there is no context. So we'll say Scotty Scheffler opens his round on hole one with a 255 yard drive. That's all you need to know. Sometimes it's okay to say nothing. Sometimes it's okay just to report what happened. Shot commentary went live at the beginning of 2025. In parallel to that, we're building the agentic system which drives our content engagement.
Shot commentary is on the top right here. If you go explore Torcast, it's our 3D data model for every golf course. You see every shot, you see the radar curve for exactly where that ball went. You'll see the direction the putts went, the distance, and then on the right you're seeing our betting profiles which are generated through our agentic system. We built a system using Bedrock to do this, which we're going to have Murali dive into with you. Betting profiles for us are content that drives engagement on non-tournament days. Thursday, Friday, Saturday, Sunday, traffic is up. Monday, Tuesday, Wednesday, no one cares. What am I going to read about unless there's something newsworthy that comes out about golf? We're not a busy website on those days.
Betting profiles allow us to push the lever on those days because we know that fantasy and gambling users get hyper-engaged in content related to statistics to make decisions. We have 156 players. There is no way I could tell a content person, your job is to write 156 very similar articles today. Does that sound rewarding? No is the answer to that. Using a recipe and agentic system to build those out, we can generate them on Monday morning at 9 o'clock. They're complete by 9:30. They're out on the website, and we're in the SEO index very quickly. These things on non-tournament days are our highest engaged content. It's not even close because the fans that care about it are looking on those days and we become the single source of truth for it.
One of the things we're really proud of on the tour is we lead SEO for our sport, which is unusual. People come to the PGA Tour first for scoring over other platforms, which is something we work really hard on in content like this to drive those results.
The Future of AI at PGA Tour: Operational Improvements and Strategic Implementation
What's next for us on our AI journey, aside from the content journey, I'll let you dive into more detail. I don't want to ruin Murali's part. For us it's the operational side. We spent a lot of work in the last year working on fan engagement ideas and betting profiles. We actually produce about 180 different types of articles, 180 articles of about 8 different types which are highly engaged things but super easy to write. At the end of a tournament we do a purse breakdown showing who won and how much money. It's a literal vertical line for search at the end of a tournament, it is the most searched thing. How much did people make? Everyone wants to know how much did Scotty Scheffler win this week, what did Rory McIlroy make this week.
We do a possible points breakdown. Given the position finishes, how is that going to affect it? These are all essentially mathematical scripted things, super easy to run through language models and create, but also highly engaging for us. Where we're changing our approach or evolving our approach, from late 2025 into 2026, is we no longer try really hard to do AI projects. We do projects to utilize AI when it makes sense.
That's been a real mindset shift I've pushed very heavily for in the tour because I think it's really easy to get excited about AI and assume it must be amazing. There's a tendency to think "I must have more AI, give me more of the AI" and let's call it agents so it sounds even cooler. The truth is that where it makes value and sense to us, we absolutely want to do it. These things do make value and improvement, but it doesn't mean it's the be all and end all.
Where we are looking at now, the value being derived is in our development processes, in our operational support processes, and in our content management and planning phases. We're asking where can we use AI tools to drive value and operational improvements. Tools like Quiro, which you'll hear a lot about this week, your Claude code, Cursor, those kinds of platforms—we're in the infancy of trying to work out which of these make a difference. To the untrained eye they look great, but the trained eye recognizes that while something looks almost good, you might have just put the secret in plain text.
The real challenge for us as we implement these tools is also making me evaluate how we hire. My standard development approach was to hire some junior developers and find some mid-tier developers to run that team. What we are seeing is that AI tools work really well when driven by more senior talent. If I give an AI coding tool to someone at a code camp, it'll generate what to them looks like great code. It looks fantastic, it's spelled well, it's camel case, it's perfect. AI agents like to succeed. They like to make you happy. They're like puppies—they want you to be happy with the result of what they produced.
Give that same tool to a senior developer and they say "that's great, but I'll probably change this pattern." That also influences how we prompt those tools. When we put those tools in the hands of a senior developer, they change their prompt. They give much more direct guidance in terms of how it works, and we see the quality that comes out. This next drive for us is really the operational improvement to my development teams, to the operational teams, to the short link teams, to every component of what we touched at the PGA Tour.
Does that mean we'll use it everywhere? No, that would be a terrific lie if I told you that. What we're getting really rigorous about is using it in the right place where we see differences being made, and also not being afraid to stop and say it doesn't work. There's a real tendency to keep going, thinking "it's five more minutes, five more minutes, it'll be fine." But the bravery to walk away is something we've gotten better at—recognizing when an idea didn't work and moving back from it.
The big one for us there, because someone's going to ask me, is image generation. It's a complete nightmare for us. We have player rights, player images, all sorts of AI and IP things. We are not having great success presently with image generation. It would be fantastic if we could, but the IP stuff makes it hard. If we look at the operational stuff inside the building, whether it's coding or AWS account analysis or spending anomalies, that type of stuff, yes, we're seeing value with those things.
Technical Deep Dive: Agentic AI Architecture Delivering 95% Cost Reduction and Billions of Page Views
I know Hannavan and Murali, who's going to dive deeper into the technology behind the agentic AI content generation. Thank you, David. My name is Murali Baktha. I'm an AWS Solutions Architect, and I've been covering PGA Tour for the last five years. David explained the use case of content generation and how content generation helps PGA Tour and brings business value. We're going to dive a little bit deeper and figure out how they implemented this and what are the things that they considered before they implemented this solution.
Generating content at scale is hard. PGA Tour generates close to 800 articles per week on different kinds of articles. They generate around 800 articles a week, and generating this content using AI is hard. There are several challenges when you generate articles with AI. The first one is that the number of articles that need to be generated is large. The second one is generating articles for live sports events is hard because LLMs typically do not have the data that is needed to generate articles.
The LLM's data is usually old data, so you need to feed data to the LLM to be able to generate these articles. You're generating these articles based on live sports real-time data of how the player played today or how the player played just a few minutes ago. You need to feed this data to the LLM to be able to generate the content because the LLM by itself does not have this data. The next thing is that when you generate content, PGA Tour being a world-class sports organization, they wanted to make sure that the brand value is not damaged.
When you generate content, you want to make sure the content is validated and has correct factual information. It also needs to comply with all the brand guidelines and style guidelines that PGA Tour has so that their brand values are maintained. In addition, one of the challenges with PGA Tour is that they generate a large number of articles. David mentioned that for betting profile articles, in a typical tournament they have 156 players, and so they generate 156 articles on each of those players. They also have betting profile summary articles, tournament preview articles before the tournament starts, and tournament recap articles after the tournament ends describing how the tournament happened.
They write articles about player recaps for each of the 156 players, detailing how each player performed in the tournament. They also write articles on round recaps. In golf, round one is played on Thursday, round two on Friday, and round four on Sunday. After each round, they write articles about how each player did. Even for players at the bottom of the leaderboard, there are fans globally who want to read about their own players. You need to write these articles to engage fans who are globally present, even though they are not in the leaderboard, to maintain fan engagement.
Not only do they write different kinds of articles on different kinds of players, but the articles can be in different formats and lengths. For a website or mobile app, they can be long-form articles. For social media posts, they might need to be a single paragraph. For app notifications, such as when Scotty Scheffler hits a hole in one, they need to generate short-form content to send to fans. There are different kinds of content that need to be generated in different lengths and different formats, which is what makes the problem challenging. Your architecture needs to accommodate all these things.
Typically, when a content request comes in to generate a particular type of content, such as a betting profile article, the first thing you want to do is research. You want to get statistical information from the PGA Tour's database through structured data available via APIs. You get structured data about how the player performed last week or last year and statistical information about the player from the tour APIs. The research request also sends a request to the data agent, which looks through the media guides. The media guides are large PDF documents containing information about where the player grew up, where they went to school, how many tournaments they have won, and all kinds of other information.
The data agent looks at the unstructured data available in the PDF document and gets the information needed to write the content. The research request gathers data from both structured data and unstructured data, and all that information is sent to create a work order. The work order specifies that you need to write a betting profile article and includes the data from research, such as statistical information from the tour's structured data as well as unstructured data from the media guides and other areas. All this information is passed on and a work order is created for the editor to make the next request.
The editor takes over the first job from the work order and asks the writer to write an article with a specific format for a specific target audience. The request is sent to the writer agent, who writes the content based on the information provided. When you look at articles on the PGA Tour website, you can see these AI-generated articles with a disclosure at the bottom stating they are generated by AI. An image is posted with each article, and that image needs to be selected from their repository.
The image needs to make sense for the specific type of article being generated and must present the specific player or PGA Tour in a positive light and angle. You don't want to show an image where the player appears angry or in a negative context. You want to select the right picture and add it with the content. This image with the written article is then sent back to the editor agent.
The editor agent reviews the article and, if it finds the article is well written and conforms to PGA Tour style guidelines and brand guidelines, sends the request back for validation. If the article is not well written or does not conform to the guidelines, the request goes back to the writer, who fixes those issues. Once the editor agent feels comfortable with all the content that has been generated, the content is sent back for validation because we want to make sure the content has the right facts in it.
You want to ensure that if, for example, Scotty Scheffler scored a birdie on the 18th hole, that content is correct. That is what validation does. You need to extract the facts from the content and use the data agent, which can give you data from the PGA Tour API. For instance, the data agent provides information that Scotty Scheffler scored a birdie on the 18th hole. That information is compared with the data extracted from the content. If they match, the verification is correct and the verification passes, and then it goes for publication.
If the data does not match, the request is sent back to the editor agent, asking them to fix the issues identified by the validation agent. Once the validation is complete, the request is sent back for publication to different channels of publication. This is the overall high-level workflow of how the agents perform.
To look at the technical architecture of how this works, this is implemented using Agent Core to run most of their agents inside the Agent Core runtime. A request comes in through the black box on the bottom right, and once the request comes in, it is returned to a DynamoDB table where a specific workload has been launched. The request then goes into a job transformer, which transforms the incoming request and puts it into an SQS queue.
As mentioned earlier, they write hundreds of articles, and generating all these articles at one time could hit some limits on token usage and other constraints. So they put all these requests in a queue. From there, they pick one request from the SQS queue, and the Lambda Agent Core invoker takes that message and makes an invocation on the Agent Core runtime. The Agent Core runtime takes the job and, as we discussed earlier, needs to make requests to the PGA Tour APIs to get data, needs to make requests to get images, and needs to make calls to an LLM on Bedrock to generate the content. PGA Tour is able to save a lot of costs by running these things in Agent Core because every time you make a call, the costs are optimized.
Specifically, when you make a request and wait for a response, you don't pay for compute resources with Agent Core runtime. The Agent Core runtime also monitors latencies and performance metrics using Agent Core observability. Once the content is generated, it's written back to S3 buckets, which then triggers a Lambda function that pushes the content back into the content ingest workflow. This is the high-level architecture that has been implemented for this content generation process.
The cost savings are remarkable. These articles are generated for 25 cents per article . On the right-hand side, you can see that each article generated using generative AI costs 25 cents, which resulted in a 95% cost reduction. Previously, whatever was spent to write these articles has now been reduced by 95%, meaning they're only spending 5% of the original cost to write these articles.
Currently, they're writing around 140 to 180 articles per week. By the end of this year, they plan to write close to 800 articles per week using agentic AI. These AI-generated articles receive billions of page views per year. To ensure image quality for the articles, they use the Nova model, which reduced costs by an additional 75% because Nova models offer superior price-to-performance value for image selection and review.
Looking at the graph, you can see that these articles have the highest views every week when the tournament happens, with peaks corresponding to tournament activity and drops when the tournament ends. Because they're using AI to generate these articles instead of humans, they can produce them within 5 to 10 minutes after the tournament ends. For example, if a game ends at 5:00 p.m., they can generate articles by 5:05 or 5:10, making them the fastest to market compared to other sports outlets.
This speed to market means they get the highest number of hits because when people search on Google, they're the only ones with that information about these articles. They're not only saving on costs but also increasing fan engagement and achieving the highest number of hits with these agentic AI-generated articles. Agentic AI can reduce your costs overall while also helping you increase fan engagement and improve performance, as demonstrated in the PGA TOUR case.
Thank you. We have 20 minutes for questions, so if you have any questions, there are two microphones at the back. Feel free to ask either me or David before we end. Thank you.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)