The Golden Whale CTO maps where machine learning can run player engagement on its own, and where operators still need to set the guardrails. 

Operators are under growing pressure to hand player engagement, retention, and bonus decisions to machine learning, but few agree on where human control should stop. In an interview with European Gaming, Golden Whale chief technology officer Thomas Kolbabek set out what an operator’s data needs to look like before its models can be trusted, the one decision he would never automate, and why he believes harm detection should sit with a separate company.

Key Takeaways 

• On players being watched: ‘We do not add surveillance; we make already existing data more useful.’
• On how the models should operate: ‘Rules and playing field must be provided, within that the models must play to win.’
• On the limits of automation: ‘Boundaries and safeguards must be set outside the models for them to operate in.’
• On responsible gambling: ‘Our models read risk signals and avoid causing harm, but that is, of course, not the same as detecting and intervening on problem gaming based on scientific research.’
• On deploying AI for the first time: ‘Automation within guardrails should follow established trust, not be deployed prematurely.’

What an operator’s data has to look like first

European Gaming (EG): Golden Whale markets itself on needing no integration, but the models still live or die on the data underneath. What does an operator’s data need to look like before Foundation’s recommendations are worth trusting?

Thomas Kolbabek: Correct, even after years of operation, we stand by the principle that operators can provide their data as is. We adapt to the format, frequency and access/delivery method that is most convenient. Data needs to represent the gaming activity and experience of a player, covering signup, payments, game play and ideally campaigns, incentives, and gamification offered and/or utilised.

We have already worked successfully with every shape of provided data, from more superficial daily aggregates down to real-time event streaming.

Whether players should be told the game is watching

EG: You describe games as sensors that read players and feed that behaviour back into the models. Most players have no idea a game is watching them that closely. Should they be told?

Kolbabek: Certainly. Regulations like GDPR and equivalent rules already mandate and enforce that users (players) are informed about what data is collected and why.

‘Digital products, including games, have always recorded data for operational and compliance purposes, including payments, game play, and other product interactions. We do not add surveillance; we make already existing data more useful.’

The one decision that never goes to the model

EG: Full Model Control points toward the system running engagement and incentive decisions on its own. What is the one decision you would never hand to the model?

Kolbabek: Boundaries and safeguards must be set outside the models for them to operate in. These boundaries are determined by company strategy, budgets and regulatory requirements.

‘I would not fully hand over budget control to models, meaning that they could propose any amount of bonus or volume of incentives to players.’

The soccer/sports analogy: rules and playing field must be provided, within that, the models must play to win.

How to tell what the model actually delivered

EG: Your case studies cite uplifts between 150% and 450%. How do you separate what the model actually delivered from what a competent team would have fixed anyway?

Kolbabek: We utilise Target/Control comparison to reliably assess impact across a configurable set of KPIs. Our analytics solution considers statistical confidence, outliers (e.g., players with large winnings) to assess impact and confidence intervals. Impact is assessed on the Campaign and Monthly level with transparent reporting.

While the Target Group is operated based on predictions and recommendations of Golden Whale, the Control Group receives the treatment strategy of the competent team (or competing tool), acting as a baseline.

Why harm detection should sit with a separate company

EG: Plenty of your competitors now sell responsible gambling AI as a core product. Is harm detection part of what Foundation does, or is that not your problem to solve?

Kolbabek: We believe responsible gaming (AI or not) is and always must be a complementary product or rule set that must be offered by a dedicated and unrelated entity.

‘A company that optimises the gaming experience (‘Red Team’) of players and limits it (‘Blue Team’) at the same time has conflicting business interests that must be avoided.’

Like in IT Security and Accounting/Auditing, we believe that these functions should not be offered by the same entity.

When the signal for engaged is also the signal for at-risk

EG: The signals that say a player is highly engaged are often the ones that say they are at risk. When the model sees both, which does it serve?

Kolbabek: Models must be kept within the boundaries set by regulation and the operator’s strategy, so that, for example, they do not provide a bonus when a player’s risk is increasing or likely to increase. Assuming a model would see both, it depends on what the model is built to predict or recommend, the ‘Target’. It will use the data points to determine the most likely outcome or recommend the next best action.

A model tasked to increase the longevity of a player will most likely not recommend a bonus for an at-risk player, as the bonus would lead to over-engagement and a subsequent block or exclusion that would be contrary to the model’s goal and a harmful player experience. 

Our models read risk signals and avoid causing harm, but that is, of course, not the same as detecting and intervening on problem gaming based on scientific research.

What operators should demand before going live

EG: Regulators have been clear that the operator, not the vendor, owns the outcome when a model gets it wrong. What should an operator demand from you before they let Foundation make live calls?

Kolbabek: What is already part of our standard process and offering:

• Ensuring that models are only trained and maintained on data that is acceptable within the applicable regulatory requirements and established responsible gaming boundaries.
• Thorough evaluation of the model performance, together with details on why certain modelling decisions (e.g., Target definition, Segment sizes, Thresholds) are made.
• Back-testing of the model on a randomly selected set of players that were not used to train the model (‘unseen’) to ensure that the model is not cheating (‘overfitting’).
• Constant evaluation of the impact on a configurable set of KPIs on both activity and monetary aspects of the gaming operation.
• Predictions are provided to the operator continuously, establishing an auditable log that can be reviewed in case outcomes deviate from the expected path.

The level of automation is ultimately decided by the operator and can be adjusted based on the associated risk/cost and potential gain/efficiency increase.

What an in-house team cannot replicate

EG: Most operators now have their own data scientists building churn and incentive models. What can Foundation do that an in-house team genuinely cannot, beyond doing it faster?

Kolbabek: Two distinct points, while keeping in mind that data of our operator partners and the associated models are strictly separated, utilising fully segregated virtual environments and state-of-the-art authentication and access control.

• Field of vision: In-house teams cannot see beyond the data and options that are within their operation; the industry-wide view of Golden Whale, paired with our in-house Loops system, allows us to refine our models in ways the single-perspective teams cannot. Our accumulated expertise in what works is the leverage that an in-house team cannot replicate. 
• Efficiency: Training models themselves is only one part of building a successful machine learning stack, especially in real-time. In-house teams have to build and operate the same tooling for a smaller set of operations, which increases the total cost of ownership compared to our platform that is developed continuously, battle-tested and has been operational for years already.

Where to start when automating for the first time

EG: For an operator about to hand real player decisions to AI for the first time, what is the single piece of advice you would give them?

Kolbabek: Start small and iterate to scale while ensuring that your team understands the goal and capability/performance of the models deployed along the way. That understanding is the basis for defining the associated guardrails (limits, human-in-the-loop, alerts).

Automation within guardrails should follow established trust, not be deployed prematurely.

About Golden Whale

Golden Whale Productions is an Austria-based data science and machine-learning company that builds real-time decisioning tools for iGaming operators and suppliers. Co-founded by chief executive Eberhard Dürrschmid, chief technology officer Thomas Kolbabek, and chief operating officer Claudia Heiling, its Foundation platform connects to an operator’s existing systems to run prediction, retention and incentive decisions without replacing the current stack. The company works with operators and game studios internationally and frames its approach around ‘Full Model Control’, the gradual handover of optimisation to machine learning within human-set boundaries. 

The post Golden Whale’s Thomas Kolbabek on AI automation: ‘The models must play to win’ appeared first on European Gaming Industry News.