A method rather than a list
Any list of emerging technologies dates quickly, and a course that provided one would be wrong within two years. What lasts is a method for assessment, which is what this lesson provides, applied to the categories currently being marketed to this sector as illustration rather than as prediction.
The assessment
Start from a problem you have. Not from the technology. The question is which existing problem this solves better than the current solution. Technologies with no clear answer are seeking a use, and the applications generated in that direction are typically weak.
Establish what the current solution costs. A technology that improves on nothing in particular is competing with a status quo that works adequately and is already paid for.
Assess regulatory readiness. In this sector this constrains adoption more often than capability. Can this be deployed within the rules of the markets served? Does it require certification? Does it process personal data in ways requiring a basis and disclosure? Does it involve automated decisions affecting individuals? A technically excellent solution that cannot be deployed in the operator's main markets is not a solution.
Calculate total cost of adoption. Integration, maintenance, staffing, certification across markets, and eventual replacement, not the licence fee.
Identify what it depends on. Many technologies require data quality, identity resolution or infrastructure the operator does not have, and the dependency is frequently larger than the technology.
Define success in advance. What would this need to demonstrate to justify adoption, expressed as a number where possible.
Establish the exit. What would cause abandonment, and what would that cost.
Ask who would own it. A technology without an accountable owner in the operating business will not be adopted regardless of its merits.
Piloting properly
Pilots are the standard mechanism and are frequently run in a way that produces no decision.
The failure is not defining success beforehand. A pilot without a threshold returns interesting findings, and interesting findings support the conclusion that further investigation is warranted, which is the default outcome and the reason organisations accumulate perpetual pilots.
A pilot that works has a specific question, a defined success threshold agreed before it starts, a bounded scope and duration, realistic conditions rather than a favourable sandbox, an owner accountable for the conclusion, and a decision at the end which may be adoption, abandonment or a specified next step with its own criteria.
The willingness to conclude that something did not work is what makes pilots valuable. Organisations that never abandon a pilot are not learning from them.
The current categories
Applied to what is currently being sold to this industry, with the caveat that this is illustration of the method rather than a settled view.
Machine learning applied to operational problems is the most established and the least discussed as a novelty, because it has been in routine use for years. Pricing models, fraud detection, recommendation, churn prediction and risk indicator identification are all standard. The assessment questions here are about data quality and evaluation rather than about capability, and the failures are generally the ones described in the measurement lesson: models optimised for the wrong objective, or deployed without guardrails.
Large language models have obvious applications in customer support, content generation and internal knowledge access. The assessment questions are about accuracy in a regulated context, since a support system giving incorrect information about bonus terms or regulatory requirements creates real exposure; about what may be disclosed, given the tipping-off and restricted-information constraints covered in the Customer Service course; about data protection where customer information is processed; and about human oversight where the output affects a customer's position. The applications with the clearest case are internal, assisting staff rather than replacing customer contact, where an error is caught before it reaches anyone.
Blockchain and cryptocurrency have been marketed to this sector for over a decade with limited adoption among regulated operators. The assessment is instructive: the problems it claims to solve, principally payment friction and provable fairness, have existing solutions in account-to-account payments and certified random number generators respectively. The regulatory readiness is poor in most regulated markets, given source of funds, sanctions screening and volatility concerns. It has genuine traction in less regulated segments, which is itself informative about where the value proposition lies.
Virtual and augmented reality have been proposed for casino environments for years. The problem being solved is unclear, since the sector's dominant trend is towards short mobile sessions rather than immersive ones, and the friction of a headset runs directly against that. The technology works; the application has not found a problem.
Provably fair mechanisms allow a player to verify an outcome was not manipulated. The problem is real in unregulated contexts and largely solved in regulated ones by certification, which is why adoption tracks regulatory status closely.
Biometric verification addresses a real problem in identity confirmation and age verification, and its constraints are regulatory and data protection rather than technical. This is a category where the assessment questions have clear answers and where adoption is proceeding accordingly.
The pattern across these is that the technologies gaining traction are those solving problems operators actually have, within the regulatory constraints they actually face. That is unremarkable and is precisely why the method matters more than the list.
Why adoption fails
Most technology failures in this sector are organisational rather than technical. The tool works and nothing changes.
No owner. Deployed by a project, handed to nobody, maintained by nobody, used by nobody.
No workflow change. The tool provides information and the existing process continues unchanged. A churn model producing predictions that no retention process consumes has produced predictions.
No trust. Users given a system they do not understand and cannot interrogate will work around it, particularly where their judgement is being displaced. This is acute in trading and compliance, where practitioners have well-founded views and will not defer to an unexplained output.
No data foundation. The recurring finding across these courses. Attribution, definitions and identity resolution are prerequisites, and organisations attempting sophisticated applications on weak foundations get sophisticated wrong answers.
No evaluation. Deployed without measurement, so nobody knows whether it helped, and it persists as a renewed licence.
The conditions that support adoption are the inverse: an accountable owner in the operating business, an explicit change to how work is done, transparency sufficient for users to trust the output, adequate data, and measurement against the criteria set before adoption.
Being late is usually cheap
A closing observation that runs against the framing of most technology discussion.
The cost of being late to a genuinely useful technology is generally modest. It arrives, it works, it is adopted a year or two after the earliest movers, and the operator captures most of the value having avoided the cost of establishing that it worked.
The cost of being early to something that does not work is the investment, the integration, the maintenance, the opportunity cost of the capacity consumed and, frequently, the difficulty of extracting the organisation from it afterwards.
This sector has abundant evidence of both. The technologies that transformed it, mobile and in-play, were adopted by everyone eventually, and the late adopters are not obviously worse off than the pioneers. The technologies that were going to transform it and did not consumed real investment from operators who moved first.
That is not an argument for never moving early. It is an argument that category enthusiasm is a poor basis for the decision, that the assessment questions in this lesson are cheap to ask, and that the honest answer to most of them is frequently that the operator has more valuable things to do with the same capacity.
Closing the course
This course has covered product under regulatory constraint, the core journeys, discovery and personalisation, responsible design, cross-vertical product, measurement and technology assessment.
The connecting theme is that product work in gambling is ordinary product work performed under unusual conditions: much of the experience is supplied by others, the product is regulated at feature level, the metrics that matter are slow and the distribution is skewed, and the design decisions carry a dimension of consequence that product work elsewhere does not.
The operators doing this well are recognisable by unglamorous things. Their journeys work. Their discovery helps. Their tools are usable. Their measurement reflects value rather than activity. And they have declined features that would have performed. That last one is the hardest and is the clearest indicator of the rest.
Automation in compliance and safer gambling
A category deserving separate treatment because it is where the most consequential technology decisions in this sector are currently being made.
Detection of gambling harm, affordability assessment, identity verification, transaction monitoring and customer risk classification are all increasingly automated, and the assessment questions differ from those for commercial applications.
Explainability matters more. A model restricting a customer's account, flagging them for intervention or declining their transaction is making a decision affecting an individual, and the operator must be able to explain the basis. Systems whose outputs cannot be interrogated are difficult to defend to a customer, an adjudicator or a regulator.
False positives and false negatives carry asymmetric costs. A harm detection model that misses cases produces the failures documented in enforcement material. One that over-flags produces intrusive interventions on customers who are fine, damaging the relationship and diluting attention across a queue too large to handle properly. Neither error is acceptable and the balance is a judgement rather than an optimisation.
Validation must be against outcomes. A model predicting harm should be evaluated against what actually happened to those customers, which requires the outcome data to exist and to be linked. Many implementations are validated against a proxy such as whether the customer subsequently self-excluded, which is informative and incomplete.
Drift is a real risk. Behaviour changes, product changes, and a model trained on historical patterns degrades. Without periodic revalidation an operator may be relying on detection that no longer works.
Automation does not transfer accountability. The obligation remains with the licence holder. A model that failed to flag a customer is not a defence, and regulators have been explicit that deploying a system does not discharge the duty to identify and act on harm.
This is the area where getting technology assessment right matters most, because the consequences of a poor implementation fall on customers rather than on a metric.
A worked assessment
To demonstrate the method, an operator is offered a system claiming to improve customer support efficiency through automated response generation.
Which problem do we have? Support contact volume, with verification and withdrawal queries dominating. Handling cost is significant and response times on live chat are longer than target.
What does the current solution cost? Agent time, with a known cost per contact.
Would this solve it better? Partially. It would reduce handling time on routine queries. It would not address the cause of the volume, which as established in the Customer Service course sits upstream in verification and withdrawal processes.
Regulatory readiness? Substantial concerns. Automated responses about account status risk disclosing restricted information, including the tipping-off case. Responses about bonus terms or regulatory requirements must be accurate. Any contact involving a safer gambling concern must reach a person immediately.
Total cost of adoption? Licensing, integration with the support platform and account systems, ongoing accuracy monitoring, and the human review capacity required to maintain it.
Dependencies? Accurate, current knowledge of terms and market-specific rules, which the operator would need to maintain regardless.
Success criteria? A measurable reduction in handling time on a defined set of routine query types, with no increase in repeat contact and no accuracy failures on regulated content.
Exit? Straightforward, since the previous process remains.
The assessment produces a narrower application than the one offered: automated assistance for agents on routine queries, with defined categories excluded entirely, rather than automated customer-facing response. That is a smaller opportunity and a defensible one, and reaching it took twenty minutes of structured questioning rather than a pilot.
Questions to ask a vendor
For practical use, the questions that produce informative answers rather than a demonstration.
Which of your customers is most similar to us, and can we speak to them? Reference customers in the same sector, at similar scale, in similar markets. Reluctance here is informative.
What does this require from us? Data, integration effort, ongoing staffing, process change. Vendors frequently understate this because it is the buyer's cost rather than theirs.
How does this behave in the markets we operate in? Regulatory constraints, certification requirements and data residency all vary, and a product built for one jurisdiction may not deploy in another.
What does the output look like when it is wrong, and how would we know? Every system has failure modes, and a vendor who cannot describe theirs has either not looked or is not saying.
How is this evaluated? Against what benchmark, on whose data, measured how. Claimed performance figures on the vendor's test set are not evidence about the operator's conditions.
What happens to our data? Where it is processed, who accesses it, what it is retained for, and whether it is used to improve the vendor's product for other customers.
What does leaving involve? Data extraction, contractual notice, and what the operator retains.
What is your roadmap and how are priorities set? For any dependency the operator will live with, the vendor's direction constrains it.
The purpose of these is not adversarial. A vendor with good answers is a better partner, and the questions surface the mismatches that otherwise appear six months into an implementation.
A note on scepticism
A final calibration, because this lesson has been consistently cautious and caution taken too far is its own failure.
Genuine transformations do happen in this sector. Mobile changed everything about how gambling products are designed and consumed. In-play betting reshaped the economics of an entire vertical. Live casino created a category that did not previously exist and now generates substantial revenue. Account-to-account payments are currently doing something similar to how deposits work in several markets.
An operator that responded to each of these with the questions in this lesson would have adopted all of them, because each had a clear answer to what problem it solved and each was demonstrably better than the alternative at the time it mattered.
The method is not designed to reject new technology. It is designed to distinguish between technologies with an answer and technologies without one, and the reason it looks sceptical is that a large proportion of what is marketed to this industry falls into the second category.
Applied honestly, it will sometimes produce the conclusion to move quickly and invest substantially. When it does, that conclusion is considerably better supported than one arrived at through category enthusiasm, and it is easier to defend internally and to fund properly. That is the practical value of asking the questions rather than the caution they frequently produce.