The product side
The previous lessons covered models that decide what to do with customers. This one covers models that make the product and run the business: the trading models that price a sportsbook, the generative tools that produce content and handle customers, the systems that help design and test games, and the operational uses in the back office. It is where the generative wave has landed hardest and where the gap between claim and practice is widest.
Trading and pricing
Sportsbook pricing was one of the first industrial uses of statistical modelling in gambling, decades before the phrase "AI" attached to it. Pre-match models estimate outcome probabilities from team and player ratings, form, schedule, injuries and market information, and the trader applies margin and judgement; in-play models update those probabilities continuously from the event's state (score, time, possession, red cards, serve), with live data from suppliers such as Sportradar now delivered from the venue to clients in less than a second, which no human could do across the many events running at once. Player-level models price props and build the correlated pricing that same-game parlays need, since the price of "player scores and team wins" is not the product of the two. Liability models forecast where money will land and recommend price moves and limits. Managed trading providers run these models for many operators at once: Sportradar, for example, describes its Managed Trading Services as trading, risk and liability management underpinned by machine-learning models; in-house trading teams run their own on the markets they choose to differentiate.
The newer developments are in data and automation rather than in model type: computer-vision tracking data from broadcasts and venues that feeds in-play models with position and event data, with suppliers such as Genius Sports offering fully automated data collection enabled by computer vision alongside human data collectors; automated trading that moves prices without a trader in the loop for lower-tier markets; and generated markets (in-play player markets and micro-markets, which Sportradar says AI and data-driven automation allow it to create in greater numbers) priced entirely by model. The regulatory interest is in fairness and integrity: bets must be settled under the operator's published rules, which in Britain must set out when a bet will be void and how errors are treated, palpable-error rules should be applied consistently, and automated pricing should not create markets that can be exploited by anyone with faster data than the operator.
Generative models in customer service
Customer support is one of the first places operators have deployed large language models, in three tiers. Assisted agents: a model drafts replies, summarises the account and suggests actions for a human who sends the message. Automated first line: a conversational agent handles routine queries (withdrawal status, verification steps, bonus terms, how limits work) with escalation to a human for anything else. And, in a few operators, automated resolution of defined cases end to end. The gains are in speed and cost; the risks are specific to gambling. A model that gives a wrong answer about a bonus term creates a dispute the operator loses; a model that handles a customer expressing distress must recognise it and escalate under the safer-gambling rules, and in Britain customer-led contact is one of the indicators of harm operators must monitor under the customer interaction rules; a model that talks a customer out of a withdrawal or into a deposit is doing what the rules forbid gambling products to do: Britain's technical standards say customers must not be given the option to cancel a withdrawal request, and must not be encouraged to keep gambling after indicating they wish to stop. The operating pattern that has emerged is a retrieval-grounded assistant (answers drawn from the operator's own terms and help content, not the model's general knowledge), with hard rules on escalation for harm, disputes and complaints, and with logs that the compliance team reviews. In the EU, from 2 August 2026 the AI Act also requires that people interacting with an AI system are told so, unless that is obvious.
Content, marketing and localisation
Operators and affiliates produce enormous volumes of text: game descriptions, market previews, promotional copy, help content, regulatory notices, and all of it in many languages. Generative models now draft much of it, with human editing, and translate and localise it at a speed and cost that has changed the economics of multi-market content. The constraints are the same as for any marketing: advertising codes apply to generated copy exactly as to written copy (a model that calls an offer "risk-free" when the customer can lose their own money risks writing a misleading ad: in Britain the advertising regulator's advice is that "risk-free" should be avoided unless the customer genuinely cannot lose their own funds), affiliate content generated at scale without adding value for users is what Google's spam policies call scaled content abuse, and sites that break those policies may rank lower or not appear in results at all, with a third-party "best casinos" page placed on an unrelated site given as an example under its site reputation policy, and personalised generated messages fall under the inducement and consent rules. Creative generation (images, video) for campaigns is in use with the same code constraints and with the added question of disclosure, which the EU's AI Act addresses from 2 August 2026: providers of generative systems must mark outputs in a machine-readable format as artificially generated, and deployers must disclose deep fakes, although an amendment adopted in July 2026 gives generative systems already on the market before that date until 2 December 2026 to comply with the marking duty.
Game design and testing
Studios use models in three ways. Mathematical design: optimisation of pay tables and feature parameters to hit target RTP, volatility and hit frequency, which is analytics that predates the current wave and remains the core of game maths. Testing: automated play of games through very large numbers of rounds to check the mathematics against the certified figures and find defects, which is how laboratories and studios have long worked, now with more sophisticated coverage; standards such as GLI-19 require a documented theoretical RTP for each game and periodic comparison with the actual RTP. And content: generated art, sound and narrative elements in production pipelines, with the intellectual-property and disclosure questions that attach. Generative models do not design games in any meaningful sense yet; they accelerate the parts of production that were already systematic.
Operations and the back office
Inside the operator, generative tools have arrived as general productivity: summarising regulatory documents, drafting policies and reports, searching internal knowledge, writing and reviewing code, producing analytics queries from questions. Compliance teams use them to monitor affiliate and marketing content at scale (a model reading every affiliate page for prohibited terms and missing disclosures), to draft suspicious activity reports from case notes, and to keep the permission matrix of market rules current. Finance uses them for reconciliation exceptions. The risk in every case is the same: a model that is wrong with confidence, in a regulated context, with a human who did not check. The governance answer is in lesson six.
Identity and verification
The identity vendors that operators integrate are themselves model businesses: document classification and forensics, face matching, liveness detection, address and database verification, all built on machine learning and now defending against generated attacks. For the operator, the relevant questions are the vendor's performance (false accept and false reject rates, by document type and market), its handling of bias across demographic groups (a well-documented weakness: the US National Institute of Standards and Technology's 2019 testing found demographic differences in false match rates across many face-recognition algorithms, and Britain's data-protection regulator tells organisations to test biometric recognition systems for bias and mitigate it), and its explanation of failures to customers who are wrongly rejected. Verification can be a decision with significant effects and, where it is taken solely by automated means, EU data-protection law (the GDPR) gives the customer the right to obtain human intervention, to express their point of view and to contest the decision.
What works and what is claimed
An honest survey of 2026 practice: sportsbook pricing, CRM scoring, fraud and anti-money-laundering models and harm monitoring are mature, valuable and, in the last case, required in Britain. Customer-service assistants, content generation and localisation are deployed and paying, with compliance guardrails that vary in quality. Automated trading of long-tail markets and computer-vision data are growing. Generated games, fully autonomous customer service and the "AI operator" of conference keynotes are not real. The operators that gain are the ones with the data platforms, the experimentation discipline and the governance to deploy models into the decision path safely; the suppliers that gain are the ones whose "AI" is a measured improvement in a specific function rather than a label.
What to take from this lesson
Trading models price pre-match and in-play markets, props and correlated parlays, with tracking data and automated pricing the current frontier and fairness the regulatory interest. Generative models handle customer service in tiers (assisted, automated first line, automated resolution), grounded in the operator's own content with hard escalation rules for harm and disputes. They draft, translate and localise content under the same advertising codes as written copy and the AI Act's disclosure duties. Studios use models for maths optimisation and automated testing rather than design. Back-office uses are general productivity with the risk of confident error in a regulated context. Identity vendors are model businesses with bias and explanation obligations. The mature, valuable uses are pricing, CRM, risk and harm monitoring; the rest is arriving, and the keynote version is not real.