The Operational Limits of AI: What Sportsbooks Must Control Before Automating Customer Decisions
Adam Lewis, CEO of AxiumAI, examines the technical, commercial and regulatory challenges of real-time sportsbook personalisation, from unreliable live data and automated content validation to proving incremental revenue and retaining human oversight.
Automating sportsbook engagement is one thing. Proving that AI generates incremental revenue while keeping customer communications accurate, compliant and consistent with safer gambling obligations is another. Delayed data feeds, incorrect pricing and changing sporting conditions can undermine automated decisions, while higher engagement figures offer little evidence of commercial value without meaningful control groups.
Adam Lewis, CEO of AxiumAI, argues that operators need to move beyond isolated AI applications towards systems capable of making decisions using live sporting and customer data. In this interview with Gaming Eminence, he examines the infrastructure needed to support real-time personalisation, the validation controls required before AI-generated content reaches customers, and how operators should measure genuine commercial uplift.
He also considers how systems designed to increase engagement should respond to signs of potential harm, and where human oversight must remain in place. The central question is not simply how much an operator can automate, but what it can prove and control before allowing AI to act on its behalf.
Gaming Eminence: AI in sportsbooks is often discussed in very broad terms. Which operator workflows are producing measurable results today, and which applications do you think remain more hype than reality?
Adam Lewis: "The workflows producing measurable results today are the ones where AI is connected directly to customer behaviour and live sporting context. That includes real-time engagement, personalised betting recommendations, intelligent bet discovery, conversational experiences and dynamic content that changes as the game changes.
The common thread is that AI is not just generating something. It is understanding the player and the moment, deciding what matters, and acting in real time.
Where the hype remains is in isolated AI features and broad claims of transformation without enough real-world proof. A chatbot on its own is not transformational. Generating more content is not transformational. The problem is not the individual capability; it is when each one operates in isolation.
The real opportunity is the intelligence layer that connects those capabilities across the business, understanding context, deciding what matters, and orchestrating which one should act, when and how.
But ultimately, that opportunity has to prove itself commercially. The test is simple: does it work live, at scale, can it create measurable incremental value, and do operators expand it once they see the results? That is the line between AI theatre and AI that actually changes the business.
Gaming Eminence: What data quality, real-time infrastructure and internal ownership need to be in place before autonomous personalisation can work reliably at scale?
Adam Lewis: "Autonomous personalisation stands or falls on how well an operator handles the reality of live sports data, which is far messier than most people admit. VAR overturns decisions, feeds drop mid-match, observations arrive late or occasionally wrong, and player intent shifts after every meaningful moment: a substitution changes expectations around next goalscorer markets, a red card reshapes how customers think about the rest of the match.
If the system understands those changes even a few minutes late, the opportunity has gone. Latency isn't a technical KPI, it's a commercial one. That's why the foundation has to be genuinely event-driven: a 24/7 streaming backbone where match and player events flow in real time, with validation happening in-stream so gaps and anomalies are caught and handled before they ever reach the decision layer (and not discovered in a report the next morning). When a feed fractures, the system needs to degrade gracefully and adapt, keeping the customer experience seamless even when the data behind it isn't.
On top of that sits the harder craft: overlaying unstructured signals, momentum shifts, sentiment, live match context onto structured behavioural profiles in a single low-latency decision layer and crystallising them into one decision in the moment. And because these systems act autonomously, compliance and responsible gambling guardrails have to sit outside the AI loop entirely - deterministic, always-on, unaffected by whatever the model decides.
And ultimately none of it holds without the right ownership model. AI can't sit solely with CRM, marketing or data science; it has to become part of the operator's core operating model, with sports specialists, data engineers, ML teams and product co-owning the execution loop. The moment those domains retreat into silos, the experience breaks at exactly the moments customers care about most - live, mid-match, in real time.
Gaming Eminence: When AI is generating and delivering real-time narratives or betting propositions, how should operators validate its relevance, accuracy, latency and failure modes before giving it customer-facing autonomy?
Adam Lewis: "Autonomy is a word that tends to create concern, but it shouldn't. It doesn't mean removing human control, it means letting AI execute millions of decisions within boundaries humans have defined. And within those boundaries, real-time AI content lives or dies on a simple distinction: creativity can be probabilistic, correctness cannot.
Give AI the right system design and it's more than capable of making the call – what to say, when to say it, how to shape it for the brand and the individual customer. What it should never be allowed to do is publish unchecked. Every output needs to clear hard, non-negotiable gates on pricing, compliance and responsible gambling before a customer ever sees it, and if anything fails to line up with the live feed, the message dies quietly in the pipeline no matter how well it's written.
The reason for that discipline is commercial as much as regulatory. Trust is the entire business in betting, and it's asymmetric. Built slowly through consistently relevant, well-timed experiences, and lost in an instant by one wrong price or a tone-deaf message mid-match. So, relevance and timing aren't nice-to-haves layered on top of accuracy; they're part of what "validated" has to mean before any system earns customer-facing autonomy.
Human oversight then has to sit where it adds the most leverage, because nobody can manually approve millions of real-time messages. For predictable in-play moments, humans pre-approve the structures and the AI fills in the context; anything unusual or low-confidence routes to a human before release, not after.
Gaming Eminence: How should an operator measure the incremental commercial impact of an AI implementation? What baselines, control groups, time periods and metrics are needed to distinguish genuine uplift from changes that would have happened anyway?
Adam Lewis: "Measuring AI impact in betting is harder than in most industries. The sporting calendar drives huge natural swings in activity, a World Cup final, an NFL opening weekend and big punters distort the picture further, with revenue so concentrated that one high roller having an unusual month can move the needle more than the AI ever could. The variance between customers is often larger than the effect you are trying to measure.
The answer is disciplined experimentation. Permanent, randomised holdout groups matched like-for-like, tested over periods long enough to span both marquee events and quiet weeks. And there's a balance to strike between micro-optimisations on softer metrics and the long-term effects that represent the ultimate goal - genuine additional value creation and incremental bottom-line NGR.
The bottom line, though, has to stay simple. Engagement metrics tell you the system is working; they don't tell you it's worth anything. What ultimately matters is sustained user activity (active days, bet frequency, retention) converting into incremental NGR: value that demonstrably wouldn't have existed without the AI. If the holdout group would have generated the same revenue anyway, all the activity in the world doesn't count. That's the number the whole measurement framework exists to defend.
Gaming Eminence: More personalised engagement can lead to increased betting activity. How should operators balance engagement and revenue objectives with safer gambling responsibilities, and which decisions should always retain meaningful human oversight?
Adam Lewis: "Behavioural AI can become one of the industry's most powerful responsible gambling capabilities. The same intelligence that understands when a customer is likely to respond positively to engagement can also identify subtle behavioural changes that suggest interaction should become more cautious.
Traditional systems often rely on fixed thresholds. Behavioural AI continuously understands patterns, trajectories and changing behaviour long before conventional KPIs begin to move.
That allows operators to intervene earlier, more consistently and with far greater precision. There should always be meaningful human oversight around customer protection policies, affordability decisions, regulatory compliance and governance.
Humans should define the principles, and AI should execute millions of individual decisions safely, consistently and transparently within those boundaries.
Responsible gambling and commercial growth are often presented as competing objectives, but increasingly, they are becoming complementary outcomes of better intelligence.
Final thoughts
The next five years won’t be defined by operators adding more AI features. They’ll be defined by AI becoming part of the operating model itself. The real shift is from fragmented systems, manual decisions and disconnected workflows to an intelligence layer that continuously understands what is happening across the business, decides what matters, and acts in real time. That means better customer decisions, better product decisions, better trading decisions, better content and more efficient operations, all connected by the same underlying intelligence.
The winners won’t be the operators with the most AI tools, but the operators with the best intelligence layer and the ability to turn that intelligence into action faster than everyone else. That is where the industry is going: AI moving beyond a tool that supports individual functions to becoming part of how the business understands, decides and acts.
