Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- iGaming AI customer service results on Intercom at a glance
- What does the operator do?
- Which helpdesk does the operator use?
- How did the operator train their AI customer service agent?
- When did the operator decide to turn on 'direct replies' to customers?
- What was the biggest thing the operator did to improve their AI agent's resolution?
- How does the operator customize their AI agent setup to work for their business?
- Teaching it the house voice
- Asking before answering
- Closing chats that have gone quiet
- What impact is the operator's AI customer service agent having now?
- Where does the operator go from here?
Created time
Sep 10, 2026 08:13 PM
Title length (<60)
Author
Last optimised
Ecomm?
Image
Publish date
Sep 10, 2026
Video
Slug
igaming-ai-customer-support-case-study
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
A high-volume online casino and sportsbook resolves 44% of 4,687 Intercom chats with AI across the last 30 days, holding 72% AI CSAT and saving 170 agent hours in that window.
A high-volume online casino and sportsbook running player support on Intercom now resolves 44% of its chats with AI. The biggest change they made was giving our agent a live line into its payments system: two payments tasks built on that connection now take 55% of what comes in. Across the most recent 30 days that has meant 4,687 AI conversations at 72% AI CSAT, with 170 hours of agent time handed back, and about a third of chats still handing over to a person.
Most AI support rollouts start with a help center and a pile of articles, and I've sat on plenty of kick-off calls that began exactly there. On a real-money betting site that only gets you so far, because the biggest slice of what arrives is not a question a help article can answer.
More often than not, they want to know where one specific deposit went, and when one specific payout lands. A general answer to that question is worse than saying nothing at all, because it reads as a brush-off.
Around those two payments tasks sit a set of plain-English guidance rules and the knowledge articles our agent wrote for itself. What follows is what they changed, in order.
iGaming AI customer service results on Intercom at a glance
Fact | A high-volume online casino and sportsbook |
Industry | iGaming: online casino and sportsbook (B2C) |
Helpdesk | Intercom (Messaging), direct replies |
Support volume | 10,000 to 20,000 tickets a month; 4,687 AI conversations in the last 30 days, rolling out progressively |
AI resolution rate | 44% over the last 30 days |
AI CSAT | 72% |
Time saved | 170 hours in the last 30 days |
Escalation rate | 32% |
Key features used | Tasks and Tools (live payments APIs), Guidance, Self-Learning, live user data, auto-close |
Knowledge sources | Intercom help-center articles, historic Intercom tickets, live player data |
Previously used | Intercom without Fin, evaluated and declined on cost before ever switching it on |
Go-live mode | Internal note replies against a test audience first, then direct replies |
What does the operator do?
The operator runs a real-money betting business: casino games and sports betting, with deposits coming in and payouts going out. The support team fields between 10,000 and 20,000 tickets every month, and our agent is rolling out progressively across that inbox.
Most of the support load here is payments, and a small set of questions gets asked thousands of times over.
Our accounts at this size all run into the same sums: huge user bases, small support teams, and revenue per contact too low to put a human on every question. Deflection is the only way the numbers work.
Which helpdesk does the operator use?
Support runs on Intercom Messaging, with our AI agent replying directly to players. When it hands over, the conversation goes to one dedicated support team inside Intercom and is shared out across its agents. The 30-day escalation rate sits at 32%.
They have never run Intercom Fin, and they told us why: they had done their own cost analysis, and at their volume, hiring support staff came out cheaper than Fin's per-resolution price. Fin was evaluated and declined before it was ever switched on.
Fin is priced at $0.99 per outcome on Intercom's own pricing page. That charge sits on top of the Intercom seats you keep paying for either way, so the only thing left to compare is the AI agents themselves.
RecruitCRM reached the same conclusion from a different starting point, rejecting Fin on cost before choosing us.
We charge per ticket instead, resolved or not. That is $0.12 a credit on Pro and $0.10 on Scale, and a typical helpdesk chat works out around $0.10. Tasks, AI Actions and AI Tagging are usage add-ons billed on top, so the bill has more than one line on it.
Our price per ticket does not climb as the AI gets better. Anyone can test that on a 30-day trial (every feature unlocked, unlimited tickets, no card).
How did the operator train their AI customer service agent?
Three sources: the help center, the ticket history, and a live feed of player data.
First, their Intercom help center, connected through the Intercom connector. That means our agent reads exactly the documentation the human team reads, and nobody had to re-author the knowledge somewhere else first.
The connector also re-reads those articles every 24 hours. When somebody edits an article, the agent has the change the next day (nobody re-uploads anything).
Second, their historic Intercom tickets, which in an inbox this payments-heavy is the source that pays off most. The house phrasing for a payout under review sits in thousands of previous replies, and training on historic tickets is what gets it into the agent.
We backfill the last 5,000 historic tickets by default and go deeper on request. If your help center is thin, or you do not have one at all, this is the way in. The AI drafts starter articles from your historic tickets.
Third, live player data through the User Data API. This one is a live connection into the operator's own systems, and it makes the two payments tasks possible at all. Without it, the best any AI agent can do on a payout question is quote the general help article on payout timings (ours included).
Those three are what is connected, and between them they cover the three kinds of question a player brings: the general one, the one the team has answered a thousand times before, and the one about this player's own money.
Only the live connection knows anything about the player in front of it.
When did the operator decide to turn on 'direct replies' to customers?
In stages, and only after the team had read a lot of drafts (which is the way we would do it too).
It went to internal note replies first, so it drafted a reply on every conversation and a human decided what to send. The team watched those drafts against a limited test audience, then switched direct replies on.
The rollout is progressive rather than all-at-once: the agent answers selected traffic, the team reads each batch before widening it, and the 4,687 conversations in the last 30 days are the traffic it has reached so far.
What was the biggest thing the operator did to improve their AI agent's resolution?
The single biggest move was to stop answering payments questions out of documentation, and let the agent look the answer up in the payments system itself.
Two of our Tools do the fetching, each one a direct call into the operator's own systems. One checks the status of a deposit, the other the status of a withdrawal. Each takes the player's transaction reference, looks it up in the operator's payments system and comes back with the real state of that transaction.
Sitting on top of each tool is a Task: the multi-step procedure that decides what to ask the player for, what to do with whatever comes back, and how to confirm it back to the player. The deposit task goes wider than a status lookup, and covers appeals and receipt checks too. Both run as usage add-ons on our side, billed per task step, and that step price covers the tool call it makes.
Over the last 30 days those two tasks handled 55% of everything that came in, so more than half the inbox now gets answered against live data rather than documentation.
Industry write-ups list deposit and withdrawal queries among the things AI can take on for an operator, usually next to a vendor's automation range. On this account those two questions are more than half the inbox, and answering them properly takes a live call into the payments system.
Self-Learning writes new knowledge articles by comparing our AI's draft to what the human agent actually sent on a handed-over ticket. The operator has a set of them live.
Two of them carry the payments load, one on deposits and one on payment reviews, and both resolve well over nine in ten of the tickets they pick up (94% and 95% in the last 30 days).
How does the operator customize their AI agent setup to work for their business?
Most of what makes our agent sound like this team is written down in plain English: a set of guidance rules, a Custom Answer for the vaguest question they get, and a short auto-close timer.
Teaching it the house voice
Guidance is a set of rules you write for the agent in ordinary words, and the operator has written them across all three types: communication style, clarification, and handover. Each entry is a short paragraph, written the way you would brief a new starter on their first morning (no developer needed at any point).
Tone accounts for most of them, so the agent talks to players the way the team always has.
The handover rules decide when the agent stops and passes the chat to a person.
Asking before answering
Picture the vaguest message a casino inbox gets: the game will not load. No game name, no error text, nothing to work with.
A guess wastes everybody's time, so the operator's clarification rules (backed by one Custom Answer written for exactly this opener) have our agent ask which game, and whether there was an error message, before it offers anything. One question up front is what gets that chat resolved.
Closing chats that have gone quiet
Chats auto-close after 15 minutes of silence on Intercom (with a short closing message), and after 60 minutes on email. The timers are set short because a chat here can go quiet mid-conversation and stay that way.
Now an open chat means somebody is waiting (the boring-but-effective kind of win).
What impact is the operator's AI customer service agent having now?
Over the last 30 days:
- 44% AI resolution rate across 4,687 conversations
- 72% AI CSAT
- 170 hours saved
- 32% escalation rate, so roughly a third of chats reach a person
- 55% of the ticket mix handled by the two live payments tasks
- 94% and 95% resolution on the two Self-Learning payments articles
Those are the numbers for the traffic our agent has been rolled out to so far; the rollout is progressive, and the team widens it batch by batch.
Resolution here is counted the way we count it everywhere: the AI handled the conversation without escalating to a human.
For context on how these rates get measured across the market, we keep an AI resolution rate benchmark study covering 195 rated deployments across 38 vendors. Every vendor defines the metric its own way, and the deployments in it are self-selected by the vendors who publish them. Read it as aggregate, directional data (and take any single vendor's number with a grain of salt).
Where does the operator go from here?
The natural next step, as we see it, is a third task on the same live connection: bonus eligibility. A player asks which bonus they can have, the agent checks their status and offers only the one that matches, and another slice of the inbox moves from documentation to live data.
AI Tagging is the quickest way to spot the next task worth building. It classifies each ticket from the text of the incoming message, so you see the mix directly (no guessing from which tasks happen to fire). Tagging runs natively inside Intercom and is billed per attribute per ticket.
After that, they keep widening the rollout batch by batch and let Self-Learning fill in the questions the AI hands over. If your inbox looks like this one, with a handful of payments questions asked thousands of times a month, the place to start is the same: a live line into the system that holds the answer.
If you'd like to see more customer stories like this one, you can browse all our case studies. There's also a high-volume prop-trading platform running the same playbook on Intercom, and our pricing if you want to model this at your own ticket volume.
