How AI Is Changing Sport, Not Replacing People: Stephen Smith on The Sports Agents

GABBY LOGAN:
Stephen, Ryan, thank you both for joining us. What might help before we delve into this is if you explain the specialities of your respective businesses, because obviously we are talking AI, but I am guessing your businesses use it in different ways. So Stephen, the areas that you use AI?

STEPHEN SMITH:
Yeah. Well, my name is Stephen Smith. I am the CEO and founder of Kitman Labs. We provide a technology and analytics platform to high performance sports teams all over the world. We help them to improve the capabilities and decision making around talent identification, talent development, how they keep their players healthy and keep them on the field, and how they keep them performing well.

GABBY LOGAN:
Okay. And Ryan, is it something similar with you? Is there a greater focus on one thing more than another?

RYAN:
Yeah, we have looked at a few different areas of AI. I think we have looked at the performance side and recruitment, but we have also looked at fan engagement as well and how we can use AI to create new tools for the fans and how it can create content and that side of things as well.

MARK CHAPMAN:
I imagine at the moment, not wishing to give either of you nervousness in your business, but I imagine this is an area that is getting increasingly flooded. Football for one sport obviously has so much money sloshing around and so much money can be made on players that clubs will have a big appetite for the best use of AI. Are you finding that the market is getting increasingly busy?

RYAN:
Yeah, I think from our side you are definitely seeing things that are not AI now being called AI. They have traditionally just been data sets or those type of things. So I think it is becoming harder to define that for people.

GABBY LOGAN:
So define that for us then. What is the difference between AI and a data set?

RYAN:
So I think what we all see now is ChatGPT and we use all of those type of models. But ultimately that is underpinned by things like machine learning and deep learning, where an AI is effectively a computer system trying to learn, reason and act on certain decisions. That is the background of where AI has come from all the way back to Alan Turing in the 50s.

Things now are relying on OpenAI’s models or relying on just data they have got and calling basic stats AI effectively.

MARK CHAPMAN:
Right. So what is the biggest advantage, Stephen, of using it? Is it down to the volume that AI can cope with?

STEPHEN SMITH:
I think efficiency and accuracy are probably two of the most exciting areas where we can make inroads right now. And I think Gabby is right, there is a lot of noise right now. There is a lot of excitement in the market, which is great. That presents a lot of opportunity.

But there have always been lots of companies that want to work within this space. There have always been lots of companies that want to sell to professional sports teams and get excited by the allure and the passion of sport.

The reality is that for the types of decisions that we are making, especially when we think of the gravity of decisions about whether you should purchase a player or the gravity of decisions about how you treat and manage a player, in terms of the impact that that can have for them, whether they take the field or not, or whether they take the field and then they get broken or not, these are career defining decisions that we are making. They are generally decisions that are hundreds of thousands if not millions of dollars in their impact.

GABBY LOGAN:
So what kind of information then is being inputted to make those decisions? Because presumably a human has to originally give that information to the AI model in the first place.

STEPHEN SMITH:
Well, yes and no. Humans are providing some insights, but there have also been devices that we have been using in this industry for 10 plus years. Whether they are wearable devices or optical devices that are capturing information about what the athletes are doing during practice and during games.

We now have every aspect of their lifestyle instrumented. How they are sleeping, how they are showing up at the training field every day, how they are recovering, their stress levels, blood and biomarkers, biomechanical information. The quantity of data that we are getting today versus 10 years ago is extremely different.

When we think of how we process that information, how we analyze that, how we interpret that, how we turn that into a decision, AI holds a huge amount of promise for us in those areas, just from a pure efficiency perspective and our ability to process that and make a decision really quickly. That is what matters in an industry like this, and AI holds the cards there.

MARK CHAPMAN:
So Stephen, just on that, we will come to the recruitment side of things in a moment, but on the injury side of things and the medical side of things, for all that AI provides, would it still need, and I apologize if these are basic questions because I am a long way from AI at the moment, but if you have all of this data and all of the access to the information, does it still need the head of performance or the head doctor or the physio or the head of sport science to interpret that for the head coach?

STEPHEN SMITH:
Yeah. One of the key tenets for us is the fact of augmentation. The idea of AI is not to replace a human. It is to augment their capabilities, to allow them to make those decisions faster and to make them more refined and to give them more weight behind that decision through the use of analytics.

So the idea is never that we are going to replace that head of performance or replace that head of medicine. That human in the loop is so crucially important. What we are going to do is give them better information, give it to them faster, and let them deploy their expertise quicker.

MARK CHAPMAN:
Is it hard to prove your case? Because if you give data and the doctor or medic who is advising a coach says, “You should not be playing this player at the weekend, we think that potentially something is going to happen because of this, this and this.” Obviously if they do not play, you never know that that would have happened. So how are you proving your case and making value for yourself?

STEPHEN SMITH:
I think it is rarely that exact decision that is happening. Unless the margins are so wide, a medic is generally not, and this is what I spent my career doing prior to this company, you are not walking into the head coach’s office saying, “Hey, you should not play this player,” unless that player is completely broken.

Generally they are building up to the critical event of a game through practice. What is happening is that we are managing the athlete every single day in practice and then we are trying to get to the game.

The decisions are happening around how we individualize their training every day. So what should they do or not do when they go into the gym? What aspects, when they are walking out onto the training field, do they maybe not do? Maybe they are not doing some of the contact elements. Maybe they are not doing some of the higher intensity components because they provide higher amounts of risk today.

The decision of whether somebody should or should not play, the players want to play, the coaches want to play them. Unless there is a huge delta, generally our job is to manage them and get them there.

GABBY LOGAN:
I suppose success would be looking at a manager’s previous season and the players under him. There are certain managers who have had a reputation for players getting broken. If you look at that manager and say, “Last season you had X amount of players out through injury but this season that has gone down by 50 percent,” you would point to your information and say, that is what we have done and that is what we have achieved.

STEPHEN SMITH:
We are also providing insights every day to say there is a risk level every day for an athlete taking the field. If normal risk level is 0.5 out of 10, that this player is at 4 out of 10 today or 5 out of 10. Coaches understand that type of terminology and nomenclature.

As long as we can point to why that is important, it is not just walking in and saying, “This guy is at risk.” This guy is at risk for these reasons, and generally the vast majority of those reasons will be modifiable. If there are things that can be done to modify that risk, that is our job.

The coach then gets to make the decision. If somebody is at 3 out of 10, the game might be so important this weekend that we take that risk. Or maybe it is not and we do not take it, we rest them and move on.

MARK CHAPMAN:
I am also wondering whether it simplifies everything for some people. I interviewed the former head of sport science at Celtic under Ange Postecoglou and he then went to Tottenham with him. By the end of that interview my head was absolutely scrambled with what was on his plate.

He was trying to help with planning the training sessions, who needed what load, who had been dropped from Thursday but was going to play on Sunday. How are they going to train Friday and Saturday?

Trying to plan training sessions for a Premier League first team who also have European football, 10 days in advance, was mind boggling because of everything that has to be fed into it for each individual before you even get to what the team needs are.

STEPHEN SMITH:
Thirty individual periodized plans that are on completely different tracks because the demands for each of the players are going to be different. What they have been doing is different and what is required of them, because of how they are going to be rotating, is different.

Then you have to take into account how the team wants to play and what the coach actually wants as well. The level of detail and sophistication that is needed to operate at that level, especially for a team that is playing European football as well as domestic football, is mind blowing.

MARK CHAPMAN:
Ryan, you focus a lot on recruitment. I imagine some of this information, in terms of injuries and how long players have been out, goes into the model. While the head of the medical department of a football club might not lose their job because they are using this information, in terms of recruitment, the people that are giving that information currently to the head of recruitment, their jobs presumably are at risk, are they not? Because you are doing their jobs for them.

RYAN:
I think it comes back to something Stephen said where it is all about combining the best humans with the best AI to come up with something that is overall better for everyone. You need the human to really trust what you are giving them.

When it comes to building, it is not always best, especially in football, to build the most complex model. It is best to build the most explainable model, something that you can really explain why it has come to certain decisions.

MARK CHAPMAN:
So where has it worked well?

RYAN:
I think what anybody can do well is look at any league and see who is scoring the most goals or who is the best defender in that league. What is really challenging is looking at all of the data around that player to see if that player is going to continue to perform well at a new club.

It is that simulation side where I find our research has been really interesting. You can give a decision maker this extra bit of information where you say, “We have simulated this player from the Bundesliga into the Premier League and we can see that they are playing against quicker defenders, stronger defenders, they are going to score 20 percent fewer goals.”

MARK CHAPMAN:
Is that Florian Wirtz?

RYAN:
No. I think the AI actually liked Wirtz. It was Werner we really liked a few years ago and obviously it was proved wrong.

MARK CHAPMAN:
So let us take Timo Werner then rather than Florian Wirtz or several strikers who have lit up their division and then come over here and not managed to do it. Your models will simulate what, say, Timo Werner’s first season over here would have expected him to do?

RYAN:
Yeah. You are looking at if that player is overperforming in their current league. You are looking at that type of data to see if they are just having an incredible season.

You look at a player like Gonçalo Ramos at Benfica, or you look at the goals a striker has scored in Portugal. Then you are saying, “What is the lowest level team he is playing against and how do we compare them to someone in England?”

The lowest teams in Portugal are more comparable to the lowest teams in the Championship in England. So you are trying to look at what percentage of the goals have come against that quality of team and then translate that across.

Then you have other things like tactics that change, language change, teammates change. Is he going to get the same quality of chances? There is so much more information than an individual scout can process themselves. We help them process all of it and provide that simulation.

MARK CHAPMAN:
Does it then, and I will turn it to Wirtz now because I cannot remember who else Chelsea signed around Werner to use this example, will your model simulate how Wirtz will do if Isak is signed, how Wirtz would do if Nunez was kept, how Wirtz would do if he replaced Gakpo on the left, how Wirtz would do if he was in a three with Szoboszlai and Mac Allister? Would it do all of those and then you hand them to a head of recruitment?

RYAN:
Yes. We break it down into looking at how he works with teammates and how he works in a certain tactical system. I spent a lot of my PhD a few years ago looking at how players work together and what qualities make a good pair of players.

A lot of data has been spent looking at how we find individuals, but actually when you come into football, compared to sports like baseball where Moneyball is really big, the way players work together is so important.

MARK CHAPMAN:
Give us an example from your PhD, the pairings that you thought worked well.

RYAN:
If you look at a player like James Ward-Prowse who is big on crosses and known for set pieces, and then you paired him up with someone like Chris Wood who is really well known for headers and winning lots of set pieces, you can see in footballing terms why those two players might link up well together.

So you are going through millions and billions of pairs of players that could be created across so many leagues to then find what are the qualities that create a really interesting pair of players and which pairs are at the central core of any team.

MARK CHAPMAN:
Although the only thing that I would say with that, without disputing that model, is that I could put James Ward-Prowse and Chris Wood together.

RYAN:
We are looking at hundreds of different features here. I am trying to give a simple example.

GABBY LOGAN:
He did his PhD in Southampton so I think James Ward-Prowse was obviously front and center of your thoughts.

RYAN:
Yeah, definitely.

MARK CHAPMAN:
Presumably it is more nuanced. If you put James Ward-Prowse and Tomas Soucek together now, does that have the legs for a West Ham midfield? Is it doing stuff that my eyes cannot see?

RYAN:
Yeah. It is trained on millions of data points that have come before, to look at how Ward-Prowse and Soucek have linked up with similar players to that person in the past and then what happens when they connect.

This is then linked to who they are playing for at the time. You have to feed in that other information to normalize for Burnley versus Arsenal. It is looking to see, when those two players link up, what is going to happen. Are they going to move the ball forward? Are they going to lose possession? Using all of these things over a number of different years to provide those simulations going forward.

MARK CHAPMAN:
Sample size is the key here, is it not? You could go back through James Ward-Prowse’s career all the way to youth team, if that evidence and data were available. Sample size is the massive advantage here because you can run all sorts of different models, different games, different players and get it into one page for a head of recruitment, quite simply.

RYAN:
Yeah, exactly. It is trying to break it down into different things that a head of recruitment can understand and process as an independent bit of advice.

If we are looking at his tactical fit to a new team, we can simulate that forward and give that as, “This player is used to playing in this formation, used to playing in these tactics, this is what your manager plays, and this is the risk you are taking in terms of how that player can adapt.”

So it becomes another bow in the quiver.

MARK CHAPMAN:
Across the summer’s transfers, as far as you know, how much of the decision making across the Premier League was using AI?

RYAN:
I think it is still coming into the game more and more. If we look at Brighton and Brentford and what they have done over recent years with their owners both coming from a betting background, they have some of the most complex data in the world because they have been motivated in their other jobs, which created their fortunes to buy their clubs.

That data is then processed in a different way by them and clearly it has worked on the pitch. Brighton has spun out their analytics platform into Jameson Analytics, which is driving Hearts, and we can see it is working at Hearts as well in Scotland.

You can really start to see the benefit of those clubs and how they are using data better than anyone else.

You have then got teams like Liverpool and Man City who have gone down the more physicist approach. They have hired Harvard professors and physicists to process tracking data. They have all of these players on the pitch that they treat as tracking particles in physics and they have built incredible models over the top which are using some AI to learn.

They are using the same systems, they have taken similar approaches and hired similar people. I think Liverpool were the first to do it and then Man City have followed suit. That is then informing more the match analytics side compared to the recruitment side like Brighton and Brentford.

GABBY LOGAN:
And does that feed into tactics, the analytics?

RYAN:
Yes, definitely. That information will then be fed to Pep Guardiola. The translation layer is really important. It needs somebody to understand what is coming out of that, being fed into the coaches, being fed into Pep, who can then execute on this information.

Sometimes you might ignore it. I am sure there are cases of that translation and that trust of the models, but you have the best people in the world building those models, so there is that element of trust at a club like City.

MARK CHAPMAN:
Stephen, do you think increasingly we will see, we have looked at recruitment, we have looked at the medical side of things, Gabby touched on tactics. We heard from Laura Harvey at the start of the pod, who has used it in the American women’s league for tactics.

Will matches be increasingly simulated through AI by coaches before an actual match to go through various tactical nuances?

STEPHEN SMITH:
I think that is already happening. They are looking at that and even some of the examples we have just discussed, those simulations are happening with different makeups, with different tactics, with different strategies in place, with different combinations of players.

Some of the more sophisticated teams are using those tools as another part of their decision making process. It is not that coaches have suddenly changed their mind and decided, “I need AI to make all my decisions for me.”

But there are organizations that want that information to be provided to a coach to help in his clinical decision making process.

MARK CHAPMAN:
But of course it will still only be a guide, Ryan, will it not? If I look at, I do not know, Tottenham vs Manchester United at the weekend, both coaches could have simulated that over and over again, looked at each other’s tactics, but AI cannot tell you that in the first 90 seconds a defender is going to let a back pass under his foot, which leads to a corner.

You cannot legislate for players having a howler or Ibrahim Konaté trying to flick the ball over his head rather than just heading it away for a goal. It cannot legislate for that.

STEPHEN SMITH:
I think it can to an extent, in terms of thinking about some of the randomness, but there are patterns of decision making and behavior that we can track through data. We can start to understand the likelihood of certain players within certain environments being more prone to an error.

Things like that we can quantify. The idea with analytics or AI is not to be 100 percent right. It is to be less wrong. It is trying to get less and less wrong and make better decisions.

GABBY LOGAN:
Presumably there are certain areas of the game, set plays for example, where you can be a lot more precise. In football compared to NFL, is there more chaos in football than there would be in American football?

STEPHEN SMITH:
I think there is plenty of chaos in that environment as well. It is snap to whistle. The idea of going from snap to whistle is very different from the beauty of being able to move from attack to defense and the open nature of football.

There is still a ton of complexity in American football introduced by the rules, the variation in players, the different roles, how both lineups match up against each other, the types of strategies and formations they play.

Their formation is not a formation that you will play for 20 minutes of a game. Every single snap could be a completely different formation, a completely different setup, a completely different lineup with different players.

That brings a ton of complexity. It is not 11 players on the field for 45 minutes. There could be 25 players that get on the field or 35 players that get on the field in that period of time.

GABBY LOGAN:
Is it just a coincidence that it feels like this season the rise of set pieces and goals from set pieces is coinciding with the rise of the use of AI?

RYAN:
I would say set pieces have more been driven by the experts who have come into the game. I think Aston Villa are a good example. They are processing data, but from a data perspective set pieces are really hard to collect data on.

If you think about how bundled up a set piece is, it is hard to break that apart and get the granularity you need when you are extracting that from video data and those type of things to look at body poses and all the tactics that go into it.

I think it has been the rise of the expert in that system. From AI, I think the way AI will now go is moving towards more expert systems and what we call agentic AI, where we are using specific AI systems for specific roles where they can go out and act on certain tasks.

In everyday life we will see this as well. If you want to book a flight, you will be able to use an AI agent to track flight prices and then book it and execute on that. You build systems that are really good at one thing. You move away from ChatGPT, which is very bad at some things.

I think we were speaking before the pod around ChatGPT not knowing who players play for.

MARK CHAPMAN:
You asked me if I use it for work and I said at the beginning of the season I had an experience where I asked ChatGPT where the key battles were going to be between Spurs and Villarreal and it suggested players who had left Spurs last season.

RYAN:
That is to do with the training data.

MARK CHAPMAN:
Worse than that it sent me to a Pizza Express on Sunday night in London that had not been there for two years.

GABBY LOGAN:
On the set piece coach, honestly, Emi Buendia scored a direct free kick on Sunday for Villa and the camera panned to Austin MacPhee, the Villa set piece coach. He is not responsible for that. He is not responsible for the direct free kick going in.

You mentioned, Ryan, the wider stuff that it can be useful for. You talked about fan engagement. So will it eventually improve fan experiences?

RYAN:
Yeah, definitely. I think it is going to bleed into every aspect of life. It is going to come into that fan experience side and give us more of an ultra personalized experience of how we watch the game.

But it covers any problem in sport. When we have been looking at this as a business, we have mapped out all of the problems that go through sports and we have got a great board of people who have led some of the biggest clubs in the country.

You are looking at problems like athlete abuse and the abuse they are receiving on social media. We have built tools that can go online and find that abuse and ultimately act on the athlete’s behalf to hide it, remove it and do all of those things to protect the athlete’s mental health.

So there are so many different applications of it.

GABBY LOGAN:
That is being used right now?

RYAN:
Yeah, that is being used. We used it with the Lions over the summer and we have used it with a few different clubs as well, to help so they can continue to have that engagement without fear of the kind of abuse that they currently have.

Social media is so important where athletes are almost forced to be on there, especially in sports like college football in the US where the only way they can get paid is through NIL deals and being online and building a profile, yet then you are opening yourself up to this abuse that they can go through.

GABBY LOGAN:
Will that translate to the wider market?

RYAN:
Yeah, definitely. This can be used for celebrities, politicians, anyone who is receiving abuse can use this to remove that and basically it is a sophisticated kind of block. It is looking at context and being able to remove that in real time so you never see it.

If there is a threat, then you are still able to react because the AI is doing that for you.

GABBY LOGAN:
Stephen, in terms of women and women sports people across all sports, I know this is a really big area to use AI to help. I am reading a book at the moment called Ultra Women, which is about ultra athletes, and something I think we all know is there has been a lack of research into women’s sports health and why injuries occur.

There are all kinds of ideas about why there is an increase in ACL injuries in women in football, but that seems to be shifting to other ideas. With this confusion around women’s medical needs in sport, how can your AI get the right information when there are not enough studies in the first place?

STEPHEN SMITH:
I am glad you brought it up because I think it presents one of the biggest risks that we deal with with AI. If the contextual information that we are dealing with is not strong enough or if the models are being trained on the wrong populations, we undermine the fairness of sport. We undermine the integrity of the games that we are playing.

If we were to build a model for ACL injuries, for example, if we were looking at every ACL injury that happens within our data set over the last number of years and we look at what biomechanical factors contribute to that, what physiological parameters contribute to it, the physical parameters, the recovery parameters, the training load parameters, we could come out with a model that helps us have a significant impact on the reduction of ACLs and some recommendations based on that.

Then if we go and deploy that and we forget that model was trained on all of our data, not just on women, and we try to deploy that for women, that does not take into account that their skeletal frame is completely different. It does not take into account that hormonal release for women athletes is completely different. It does not take into account where in their menstruation cycle they are right now.

For all of those reasons, we would provide them with inaccurate insights and we may provide them with insights that are actually more dangerous for their current state.

I think that is a really important debate to be had in how we think about building these models for specific populations and groups. That is not just women, but also youth athletes or athletes of different ethnicities. There are lots of different considerations that need to be given to how we deal with data, how we build models and how well we communicate those limitations to practitioners who will leverage them.

The last thing that we want to do with any of this is to undermine decision making. The idea is to improve it.

But I do think with women athletes in particular, it probably has the opportunity to accelerate research. You are right, there is a huge lack of research right now, but there is more instrumentation in women’s sports as well. The amount of data being collected on their physicality, training load, recovery, even on menstruation itself, is increasing.

We are getting more teams that are focused on collecting information about the menstrual cycle, but we also know there is a lack of understanding of what to do with it. If AI can start to pull those pieces together and help us to develop insights and research much faster than we could in a traditional research setting, maybe that allows us to learn much quicker in women’s sports and have better application faster.

MARK CHAPMAN:
Just a final one from me. It is a bit crass. Is it expensive? Will it mean that the haves will increase the gaps over the have-nots?

STEPHEN SMITH:
I think it democratizes the industry. I think the opposite. Right now, and I think Ryan mentioned this already, who did he mention? Liverpool, Man City, clubs like that, with Brighton etc. They are spending money in certain areas, but it is all the big clubs that got mentioned. Why? Because they have more money and they have hired physicists, they have hired a team of graduates from Harvard.

AI is actually democratizing the entire industry. I think this gives every player a chance. I think this levels the playing field and it means all this computational power that has been built by huge companies that we can sit on top of, we can access and leverage that tooling, which means we can make it so much more affordable than hiring an army of data scientists and analysts to come in.

You allow the underdog a chance. Sport is one of the most beautiful places for it, because that is the beauty of sport, that everybody should have a fair crack.

MARK CHAPMAN:
Do you agree with that?

RYAN:
Yeah, definitely. I think we will start to see it bleed down the game and come into lower leagues as well. Even this week Reading appointed, I think, the first Head of AI at any professional club. You are starting to see that go into League One.

There are plenty of clubs in League One and League Two starting to do things smarter than before. I think talent is expensive because you are competing with Meta and Google to get the best AI talent. But as you see more of that come into the game and you have got those leading examples of how it is being used at the top, you will start to be able to use these models that are always getting cheaper.

In the last year it has come down exponentially, with DeepSeek and Google having their own models and OpenAI obviously. You have all of these companies competing to train models, which is great for us as consumers and in sport because you are starting to see AI prices come down.

GABBY LOGAN:
And then soon, Mark, it will be your Sunday league team. They will have their own AI.

MARK CHAPMAN:
Yeah, why not? Maybe the team in Wither with all the ex pros. Maybe they will be the first vets team. I am still not sure it is going to get me on social media though.

It is absolutely fascinating. I think you have educated both of us and we could probably sit here for another hour finding out more, but I am sure it will be something that comes up again and again over the years to come because it has clearly proliferated the upper tiers of sport and it is only a matter of time before it becomes ubiquitous.

The Sports Agentswith Gabby Logan and Mark Chapman.