DreamCasino · iGaming · Generative AI, Trustworthy AI · Published 2026-11-06

An AI player support assistant that answers in seconds, in every player's language

AXIOMIO built a multilingual AI support assistant into DreamCasino's AWS platform on Amazon Bedrock. About 70% of player questions are now resolved without an agent, with a p95 response time under 10 seconds at any hour.

Built on AWS with: gpt-oss-120b, Qwen3 32B, Cohere Embed Multilingual v3

Results at a glance

• Under 10s p95 time to first response, at any hour
• About 70% of player questions resolved without an agent
• About 80% less agent time per escalated case
• Under 1 hour for new bonus terms to reach answers

About DreamCasino

Norway-headquartered technology company providing digital services to casino and sports betting operators, expanding into Africa, the United States and Brazil. Location: Norway, serving Europe, Brazil, Africa and the US.

Overview

DreamCasino is a technology company headquartered in Norway that provides digital services to casino and sports betting operators. Founded in December 2020, it runs a microservices platform on AWS and is expanding from its European base into Africa, the United States and Brazil. AXIOMIO has managed DreamCasino's AWS environment since December 2022, after moving the platform onto Amazon ECS on AWS Fargate and Amazon MSK.

Growth into new markets brought a support problem with it. Players ask why a bonus cannot be withdrawn yet, where a withdrawal is, or why an identity document was rejected, in many languages and at every hour, and each operator has its own bonus terms and payment rules. Human agents were answering the same questions over and over, and overnight questions waited until morning.

AXIOMIO built an AI player support assistant into the platform and runs it in each AWS Region that serves DreamCasino's markets. On Amazon Bedrock, gpt-oss-120b answers players in their own language, grounded in each operator's own terms through Amazon Bedrock Knowledge Bases, and checks the player's real account status through read-only platform APIs before it answers. Bedrock Guardrails check every reply, and anything the assistant should not settle goes to a human agent with a summary written by Qwen3 32B.

The challenge

1. The same questions, answered by hand. Most contacts were about bonus wagering progress, withdrawal status, identity checks and deposit methods. Each one still took an agent's time to read, look up and answer.
2. Slow answers outside office hours. A question asked at 2 a.m. waited hours for a reply, and a player waiting on a withdrawal is the player most likely to complain or leave.
3. Many languages, many markets. Expansion added players writing in Portuguese and other languages to a Norwegian and English base. Native speakers for every market and every shift did not scale.
4. Every operator has its own rules. Each operator sets its own bonus terms, wagering requirements, payment limits and identity rules, and agents had to know which applied to which player.
5. Answers that needed the player's own account. "Why can't I withdraw?" depends on that player's wagering progress, document status or pending transactions. Agents switched between back-office screens to find it.
6. Responsible gambling in every conversation. Operators must never encourage a player to gamble more than they intend, and must point players who show signs of harm to the right tools. Any automated answer had to meet that standard every time.

The solution

1. Built into the platform DreamCasino already runs. The assistant runs as its own service on Amazon ECS with AWS Fargate beside the existing microservices, in private subnets across two Availability Zones, behind Amazon CloudFront and AWS WAF. Fargate adds tasks as chat volume rises in the evening peak.
2. Deployed close to players in every market. A complete stack runs in each AWS Region that serves DreamCasino's markets. Route 53 latency-based routing sends each chat to the nearest Region, and S3 Cross-Region Replication keeps operator documents in every Region for failover.
3. Answers in the player's own language (Generative AI). gpt-oss-120b on Amazon Bedrock understands the question in whatever language it was written and replies in that language. AXIOMIO tested candidate models on DreamCasino's own player questions in Portuguese, Norwegian and English, and chose it for accuracy, adherence to the responsible gambling rules and low cost per conversation.
4. Grounded in each operator's own terms. Bonus terms, payment rules, identity requirements and responsible gambling policies are indexed by Amazon Bedrock Knowledge Bases with Cohere Embed Multilingual v3, so a question in Portuguese finds the right passage even when the terms are in English. Every answer comes only from that player's operator, and new campaign terms are live the same hour.
5. Checks the real account before answering. The assistant calls read-only platform APIs for that player only: wagering progress, withdrawal status, identity check results. It reads the account; it can never change a balance, a bonus or a limit. Live deposit, bet and withdrawal events come from Amazon MSK.
6. Guardrails on every reply (Trustworthy AI). Amazon Bedrock Guardrails check every reply before the player sees it: no betting tips, no promised winnings, no encouragement to deposit or play more. Players showing signs of distress are pointed to deposit limits, time-out and self-exclusion tools and offered a human agent straight away.
7. A clean handover to a human agent. When a player asks for a person, the conversation passes to the operator's helpdesk with a short summary from Qwen3 32B: who the player is, what they asked, what was checked and what they were told. The agent starts from the facts instead of the transcript.
8. Managed like the rest of the platform. Changes ship through AWS CodePipeline and CodeBuild. Amazon CloudWatch tracks p95 response time per Region, escalation rates and errors, and AXIOMIO's managed operations team receives alarms and security findings through Amazon SNS.

Architecture

1. A player opens the chat on their operator's site. The request passes through Amazon CloudFront and AWS WAF, and Route 53 latency-based routing sends it to the nearest Regional deployment, where the Application Load Balancer passes it to the assistant service on ECS with Fargate. The service loads the conversation from Amazon ElastiCache and identifies the operator from the signed session.
2. The question goes to gpt-oss-120b on Amazon Bedrock through VPC interface endpoints. Bedrock Knowledge Bases retrieve the passages from that operator's documents in Amazon S3, and the service adds the player's account status from the read-only APIs and recent events from Amazon MSK. The model writes the reply, Bedrock Guardrails check it, and it streams back to the player.
3. When a person is needed, the conversation goes to the operator's helpdesk with a Qwen3 32B summary. IAM task roles, AWS KMS and AWS Secrets Manager protect access and data, and CloudWatch, CloudTrail, GuardDuty and Security Hub report to AXIOMIO's operations team through Amazon SNS.

AI and AWS services: what each one does here

Supporting AWS services: Amazon VPC, Amazon CloudFront, AWS WAF, Application Load Balancer, S3 Cross-Region Replication, AWS CodePipeline, AWS CodeBuild, Amazon ECR, AWS IAM, AWS KMS, AWS Secrets Manager, Amazon CloudWatch, Amazon SNS, AWS CloudTrail, Amazon GuardDuty, AWS Security Hub.

Other tools: Operator helpdesk systems and GitHub.

In action

Bonus wagering: Answering a bonus question at 2 a.m.

Scenario: A player in Brazil writes in Portuguese at 2 a.m. asking why the winnings from a welcome bonus cannot be withdrawn.

The assistant checked the account through the read-only API and found the bonus 72% wagered. Knowledge Bases retrieved the operator's welcome bonus terms, including which games count fully and which count at 10%. The assistant replied in Portuguese within seconds: how much wagering remained, which games count, and that the player could instead withdraw the deposit balance by giving up the bonus. Guardrails confirmed the reply set out the options without urging the player to keep playing. No agent was needed.

Withdrawal on hold: Handing a frustrated player to an agent

Scenario: A player in Norway asks why a withdrawal has been pending for two days.

The assistant found the withdrawal held for an identity check, with the proof of address rejected for being more than three months old, explained this in Norwegian and listed the accepted documents. The player asked for a person, so the conversation passed to the helpdesk with Qwen3 32B's summary. The agent approved a newly uploaded document and confirmed the payout in under three minutes.

Results: before and after

Four indicators show what changed once the assistant went live.

• Every market in its own language. Players in Brazil, Norway and DreamCasino's other markets are answered in their own language without a native-speaking agent on every shift.
• Answers specific to the player. Because the assistant checks the account first, players get their own wagering progress or withdrawal status, not a general policy paragraph.
• Responsible gambling built into every reply. Guardrails keep every reply free of tips, promises and encouragement to play, and players who show signs of harm are offered safer gambling tools and a person.
• One assistant for every operator. Each operator's players are answered from that operator's own terms, so DreamCasino offers the assistant to every operator from one service.

Security and responsible AI

• Private network. The assistant, account APIs and data stores sit in private subnets; only the load balancer, behind CloudFront and AWS WAF, faces the internet, and Bedrock and S3 calls go through VPC endpoints.
• Conversations stay in their Region. Both models run on demand in the same Region as the assistant, so questions, answers and handover summaries are processed in that Region.
• Read-only account access. The assistant can read wagering progress, withdrawal and document status. It has no permission to change balances, bonuses, limits or documents.
• Each player sees only their own data. Every account lookup is bound to the player in the signed session, and every knowledge base query is filtered to that player's operator.
• Encryption and least privilege. TLS in transit, AWS KMS at rest, helpdesk credentials in Secrets Manager, and IAM task roles scoped to what each service needs.
• 24/7 threat detection and audit. GuardDuty, Security Hub and CloudTrail report every finding and alarm to AXIOMIO's managed operations team.

Source: DreamCasino AI Player Support Assistant case study, November 6, 2026. Figures as reported in the case study.

Work with AXIOMIO

AXIOMIO is an AWS Advanced Tier Partner, AWS MSP and Level 1 MSSP that builds and runs production AI on AWS. Generative AI services on AWS · RAG development services · Talk to an AI engineer