- Support QA: coverage from about 2% to 25% of conversations and evaluator agreement with the QA lead from 66% to 91%, on a low-cost model, by making the model fill a structured schema before scoring.
- Responsible gambling: alerts reduced from about 200 a day to about 25 while catching more real cases, by combining conversation analysis with behavioural signals and trends.
- Bonus abuse: organised rings found by clustering play behaviour, without player PII and without relying on device fingerprints.
- Retention and Tier-1 support: bonus budget moved away from players who would have returned anyway; the routine Tier-1 majority resolved by AI with CSAT held or improved.
- Every number is a claim about a definition and a baseline. Ask what was measured, against what, on whose data and over what period — of us and of anyone else.
Most iGaming case studies follow the same template: an anonymous operator, a headline percentage, a quote. What they rarely say is what the percentage measured, what it was compared with, and whether it would mean the same thing on your book. That makes them hard to use, even when the result is real.
This article presents five AI systems built in production iGaming, in a different format. For each one: the problem, the mechanism, the published result, what that number does and does not measure, and what would transfer to another operator. Four were built inside a large European operator running six brands in 2024–2025; one comes from a 2026 operator deployment. Operators are anonymised, and we quote operator-specific results only after validating them against that operator's own baseline. The short versions are on our case patterns page.
The five patterns at a glance
| Area | Published result | Mechanism | Status today |
|---|---|---|---|
| Support QA | Coverage ~2% → 25%; evaluator accuracy 66% → 91% | Schema-guided scoring on a low-cost model | Part of PlayerDesk |
| Responsible gambling | Alerts ~200/day → ~25/day, more real cases caught | Conversation analysis fused with behavioural signals | Built, not yet for sale |
| Bonus abuse | Organised rings caught without PII or fingerprinting | Behavioural clustering on play patterns | Built, not yet for sale |
| Retention | Reactivation budget moved away from organic returners | Uplift modelling instead of churn scores | Built, not yet for sale |
| Tier-1 support | Routine Tier-1 majority resolved by AI; CSAT held or improved | Multi-agent decomposition | PlayerDesk |
1. Support QA: from 2% coverage to 25%
The problem. The QA team reviewed about 2% of support conversations by hand. That was enough for coaching individual agents, but not for seeing problems in a particular topic, language or shift. An early automated grader existed, but its scores agreed with the QA lead's only 66% of the time — too often wrong to act on.
The mechanism. Instead of asking the model for a score in free form, it was required to fill a structured schema first — what the player asked, what the agent did, which checks applied, where the evidence was — and only then score. This approach, known as Schema-Guided Reasoning (SGR), sharply reduced invented reasoning and allowed a cheap model to score reliably.
The result. Evaluator accuracy rose from 66% to 91%, and QA coverage from about 2% to 25% of conversations, both on a low-cost model.
What the numbers measure. The 91% is agreement between the automated grader and the operator's QA lead on hand-scored conversations. It says how far the grader can be trusted; it says nothing on its own about how good the agents were. The 25% was the coverage reached in that system; once a grader is calibrated and trusted, coverage becomes a question of running cost rather than reviewer time, which is why PlayerDesk is designed to score every conversation.
What transfers. The method: hand-label a baseline with your QA lead, calibrate the grader against it, report agreement per dimension before trusting any score. The rubric and calibration procedure are in iGaming support quality control software and the QA scorecard template.
2. Responsible gambling: fewer alerts, more real cases
The problem. The RG team received around 200 alerts a day from rule-based triggers. Most were false positives. A team buried in alerts reviews each one less carefully, and the real cases — the ones that lead to harm and to regulatory findings — get the same thirty seconds as the noise.
The mechanism. Language-model analysis of player communications, fused with behavioural signals — deposit velocity, session patterns, loss-chasing patterns — and tracked as trends over time rather than single events. Schema-guided reasoning kept each alert's reasoning structured and auditable.
The result. Alert volume fell from about 200 a day to about 25, with more real cases caught, not fewer.
What the number measures. "Fewer alerts" alone would be easy and dangerous: any system can raise fewer alerts by raising its threshold. The claim that matters is the second half — more confirmed cases among fewer alerts — which is a statement about precision and recall against the RG team's own case outcomes. Recall in RG has to be checked deliberately, with oversampled evaluation sets, because missed cases do not announce themselves.
What transfers. Layered detection tuned for recall, and a detect-hold-handoff rule in support. The support side is described in responsible gambling signals in support chats; the full system is described on the SafePlay page.
3. Bonus abuse: the rings the rule engine couldn't see
The problem. Rule-based promo protection — IP blocks, device fingerprints, KYC checks — caught clumsy abuse, but organised groups using residential proxies, antidetect browsers and real identities passed through. Each account looked clean on its own.
The mechanism. Behavioural clustering on play data: bet rhythm, session shape, deposit cadence and withdrawal timing. Accounts run by the same group share a behavioural signature even when their surface attributes differ. The fraud team received a ranked review queue — cluster, accounts, evidence — rather than a pile of individual flags.
The result. Organised bonus-abuse rings were caught without touching player PII and without depending on fingerprinting.
What the result measures. This one is qualitative by design: the value of a ring detector depends on the operator's promo mix and how much abuse is present, so a single savings figure would not transfer. The useful measures on a new book are confirmed cases per review hour and bonus cost per acquisition source before and after. More context is in bonus abuse in online casinos.
4. Retention: stop paying players who'd return anyway
The problem. Reactivation bonuses were sent to players with a high churn score. Many of those players would have come back without any bonus, so part of the budget subsidised organic returns.
The mechanism. Causal uplift modelling: score each player on the difference between the probability of a deposit with the bonus and without it, rather than on churn risk alone, and validate with a hold-out group every campaign cycle.
The result. Reactivation budget was redirected away from players who would have returned anyway; the CRM team stopped subsidising organic comebacks.
What the result measures. Uplift claims are only meaningful against a hold-out. Without one, any campaign looks successful because some players were coming back regardless. Ask for hold-out design before believing any retention figure, including ours. The design is on the Retention Engine page; how it fits with support signals is in casino CRM and support signals; the cheaper fixes that come first are in how to reduce churn in iGaming.
Twelve questions to ask about any case study — ours included.
A free PDF on buying AI support in iGaming: measurement, economics, data and regulation, and operational reality, with a one-page vendor scorecard.
5. Tier-1 support without players noticing
The problem. A support desk staffed for peak around the clock, with most of the cost going on routine Tier-1 contacts — payments status, verification, bonus questions — while VIP and complex cases competed for the same agents.
The mechanism. Instead of one large chatbot, a set of narrow agents: triage, knowledge, action, escalation and real-time QA. Each one does one job, predictably and auditably. Account actions go through rule-based validators; VIPs, RG signals, complaints and fraud go to humans by rule.
The result. The routine Tier-1 majority was resolved by AI, human cost was concentrated on VIP and complex cases, and CSAT held or improved.
What the result measures. "Majority resolved" depends on the definition of resolved and on the denominator. When you assess any such claim, ask whether a contact counts as resolved when it was closed without a human, or only when the player also did not come back about the same issue, and whether resolved contacts were scored on the same rubric as human agents. CSAT here is a guardrail — evidence that players did not notice a drop — not proof of correctness. Exact automation shares are quoted per operator after validation, never as a general promise. This is the system we sell today as PlayerDesk.
How to read any iGaming case study
A single vendor's case study can contain several different numbers for what sounds like one result: share of contacts touched by the AI, share closed without a human, share closed without a repeat contact, and share that passed a quality rubric. All can be honest; they are not interchangeable. Before comparing any two case studies, ask:
| Question | Why it matters |
|---|---|
| What is the exact definition? | "Resolved", "automated" and "handled" are different measurements |
| What was the baseline? | An improvement over a weak baseline says little about the system |
| What is the denominator? | All contacts, or only the topics the AI was allowed to handle? |
| Whose data, over what period? | A curated test set and a quiet month are not live traffic |
| Who scored quality, against what rubric? | A vendor scoring its own output with its own rubric is not independent |
| What happened to the sensitive cases? | RG, VIP and complaint handling should be reported, not averaged away |
Our own answer to these questions is the measurement methodology: baseline on your own agents, shadow mode on live traffic, topic-by-topic rollout with automatic rollback, and a labelled evaluation set that stays with you.
A number without a definition and a baseline is a claim, not a result.
What transfers to another operator
The mechanisms transfer: schema-guided grading, layered RG detection, behavioural clustering, uplift with hold-outs, narrow agents with hard routing rules. The numbers do not transfer automatically. They depend on topic mix, languages, current QA coverage, alert rules, promo design and data quality. That is why every engagement starts by measuring your own baseline, and why we publish patterns rather than promises. If you want to see what the baseline looks like on your queue before any commitment, the Support Load Report is a free first step.
The short version
Five production systems, five results: QA coverage from about 2% to 25% with grader agreement from 66% to 91%; RG alerts from about 200 to about 25 a day with more real cases caught; bonus-abuse rings found without PII; retention budget moved off organic returners; and the routine Tier-1 majority resolved by AI with CSAT held. Each number is a claim about a definition, a baseline and a dataset, and should be read that way — ours and everyone else's. For the full picture of what an AI support agent does in iGaming, start with AI player support for iGaming operators.
Frequently asked questions
Is there an iGaming AI case study with real numbers?
Yes. The five patterns in this article come from production systems built inside a large European iGaming operator running six brands in 2024-2025 and from a 2026 operator deployment, with published results such as QA coverage rising from about 2% to 25% of conversations and responsible-gambling alerts falling from about 200 to about 25 a day while catching more real cases. Each result is given with its mechanism and its limits.
What does '66% to 91% evaluator accuracy' mean?
It is the agreement between the automated QA grader and the operator's own QA lead on a hand-scored set of real conversations, before and after the scoring method was changed to make the model fill a structured schema before giving a score. It measures how far the grader can be trusted, not how good the support agents or the AI agent were.
Why are the operators in the case study anonymised?
Operators rarely allow vendors to publish their names or internal figures, and some of the systems involve responsible-gambling and fraud data. The results are therefore published as anonymised patterns with the mechanism and the origin stated, and operator-specific results are quoted only after validation against that operator's own baseline.
How should I evaluate an iGaming vendor's case study?
Ask for the definition of the headline metric, the baseline it was compared with, whose data it was measured on, the time period, the denominator (all contacts, or only those the AI handled), and who scored quality. If a case study cannot answer these, the number is not yet comparable with anything else.
Will these results transfer to my operation?
The mechanisms transfer; the numbers have to be re-measured. Results depend on the topic mix, languages, current QA coverage, alerting rules and data quality of each operator, which is why any deployment starts with a baseline on the operator's own agents and data before results are claimed.
Last reviewed 17 November 2026. Figures marked as industry ranges are ranges, not Arctura claims; operator-specific numbers are only quoted after validation on the operator's own baseline.


