When the algorithm knows the player

Personalisation promises gambling companies a better understanding of their customers. Yet the central challenge is whether they will use that knowledge to recognise when to leave someone alone.

Two sportsbook customers stop betting. One is waiting for the football season to restart. The other has an unresolved complaint and is considering leaving. To a system looking only at declining activity, they may appear much the same: two customers to win back.

Send both a bonus and the campaign may look efficient. It may also misunderstand both people. The first might have returned anyway. The second was waiting for an answer.

This hypothetical example captures a problem at the heart of personalisation in online gambling. Knowing that someone has changed their behaviour is different from knowing why. Predicting what might bring them back raises a further question: should the operator try?

A player’s account history still shows only part of the picture. It records what a player does with that operator, without explaining everything affecting their finances, intentions or choices elsewhere.

An algorithm can identify patterns within that history. It cannot turn an incomplete record into a complete understanding of a person. Yet its estimates can influence which games someone sees, when a message arrives and whether an incentive follows a period of inactivity.

These decisions have commercial value. They also carry consequences for people whose circumstances the operator may only partly understand. The responsibility lies in deciding what the evidence justifies before a prediction becomes an offer.

A prediction is not an explanation

Personalisation can certainly improve customer experience. Someone who mainly follows football may appreciate finding it quickly, while a customer with an unresolved complaint benefits from having that issue recognised before another promotional message arrives. Used thoughtfully, information reduces irrelevant contact.

But commercial predictions answer narrower questions than their apparent precision suggests. A model may estimate that a customer is likely to stop playing. That does not establish that a bonus will persuade them to stay.

If a player returns after receiving an incentive, that alone does not show whether the offer changed their decision or they would have returned anyway.

Nor would evidence that an incentive works settle whether sending it was appropriate. The ability to influence behaviour creates a reason to examine the consequences of doing so.

A high predicted player value does not show what that player can afford to spend, and a churn-risk score does not, by itself, establish gambling harm.

Language matters here. Describe someone simply as “high risk” and the meaning can shift between departments. Marketing may understand a customer at risk of leaving; a protection team may understand someone at risk of harm. A label that conceals its purpose can travel further than the evidence behind it.

The recommendations themselves can also influence what a model later reads as preference. A game given a prominent position may receive more play partly because it is easier to find.

If those automated choices then shape future recommendations, the system risks treating the effects of its own design as independent customer preference. Truly understanding behaviour requires examining the options presented as well as the choices made.

When the warning fails to stop the offer

Now consider a more serious hypothetical case. During a support exchange, an operator identifies strong indicators of gambling harm. A marketing restriction is recorded. Later, an automated campaign sends the customer a personalized bonus.

The prediction itself may well have been accurate. The failure occurs because the company acted on that prediction despite having identified a clear reason to stop.

In Great Britain, the Gambling Commission’s requirements for remote licensees within scope are explicit: strong indicators of harm must lead to the prevention of direct and targeted marketing and the take-up of new bonus offers. A restriction that exists in an account record but fails to stop the promotion has not achieved its purpose.

It is possible for several systems to perform their assigned tasks while the organisation fails the customer. A protection tool records the warning. A model selects the recipient. Campaign software delivers the message. The crucial instruction is lost between them.

That is a management failure to investigate, even if its immediate cause is technical. Who decided which information the campaign would check? Who ensured that protection restrictions took precedence? Who could suspend delivery when information was missing or delayed?

Before an automated offer reaches the player, people have chosen which information the system can use, what it should predict and which actions it may trigger.

Consequently, responsibility cannot be hidden behind the phrase “the algorithm decided.”

The operator remains responsible when a third party supplies the software. The Gambling Commission’s third-party provision holds licensees responsible for contractors’ actions connected with their licensed activities. Buying a service does not transfer that responsibility to its supplier.

Operators therefore need more than assurances about predictive accuracy. They need access to the records required to investigate an outcome, a clear understanding of the service’s limitations and a workable means of stopping an inappropriate action. These are questions to settle when buying the technology.

‍

‍‍

Human judgement needs authority

Requiring a person to approve every recommendation would do little to solve this problem. Staff could spend their time endorsing routine decisions while more serious failures remained buried in the system.

The level of scrutiny should reflect the consequences. Changing the order of games in a lobby calls for oversight of how recommendations work and what effects they produce. Sending a promotional incentive requires checks on permissions, eligibility and protection restrictions before delivery. A consequential restriction on an account requires evidence, a route to review and someone able to resolve an error.

Human involvement must also be more than a reassuring label. A reviewer needs enough information to question the decision and enough authority to change it. If the only practical option is to accept the system’s conclusion, the review provides little protection.

There are circumstances in which automation is itself an essential safeguard. The Gambling Commission requires relevant remote licensees to take timely automated action where strong indicators of harm arise, alongside manual review of how those processes operated in each customer’s case and an opportunity to contest automated decisions. Waiting for a person to approve every protective action could introduce a dangerous delay.

The challenge is to combine prompt protection with meaningful scrutiny. A temporary precaution needs an owner and timely follow-up. A corrected decision needs to reach every connected system that might otherwise repeat the mistake.

‍

An individual review can also expose a broader problem. Repeated reversals may point to unclear policy, inadequate training or weaknesses in the model itself.

Management should want to know about those reversals. A dashboard showing revenue and engagement can describe commercial performance while revealing little about whether the system is making defensible decisions. How often do protection restrictions fail? How long do disputed cases wait? Are the same errors recurring?

These measures help establish what the technology is actually doing to the people using the service.

The right to understand

For a customer facing a consequential automated decision, an explanation must make sense outside the company that produced it.

In its 2025 Dun & Bradstreet Austria judgment, concerning automated credit assessment, the Court of Justice of the EU said the explanation must allow the individual to understand how their personal data were used in reaching the decision. Simply providing an algorithm would not suffice. The case concerns GDPR rights around significant automated decisions; it does not place every personalised game recommendation in the same legal category.

The practical question extends beyond the legal threshold. Can the company explain the evidence behind an action, recognise what it does not know and correct an outcome when that evidence proves wrong?

That requires boundaries around information as well as decisions. Data collected to investigate a protection concern should not automatically become available for promotional targeting. A company’s ability to connect its systems does not establish that every possible connection is appropriate.

It also requires accepting that a customer may choose to leave. Declining activity can signal dissatisfaction, a seasonal interest, financial pressure or a simple change of mind. An operator cannot assume that every departure is a problem it should use technology to reverse.

Two customers, two distinct decisions

Returning to the two customers described at the start: one may return when the football season resumes, while the other may leave even after their complaint is resolved. Neither outcome alone reveals whether the operator acted responsibly.

Instead, the true test is whether the company evaluated the available information with care, acknowledged the limits of its inferences, and allowed player protection considerations to override commercial incentives.

A gambling operator that genuinely understands its players will recognise when an offer is inappropriate. Ultimately, the clearest proof that personalization works lies in the messages an operator chooses not to send.

‍

‍