What AI agents are — and how they differ from chatbots
Most “AI” in iGaming so far has been one of two things: a chatbot that answers from a script and hands over to a human, or a model that scores something — a churn risk, a fraud probability, a game recommendation — and leaves the action to a person.
An AI agent closes the loop. It understands the request, goes into the systems your team already uses, finds out what actually happened, takes the action your rules allow and comes back with a result. A person is involved only where a decision goes beyond the rules: money outside a limit, an unusual case, anything you mark as a human decision.
| Chatbot | Scoring model | AI agent | |
|---|---|---|---|
| What it does | Answers from a script or knowledge base | Predicts a probability | Checks systems, acts, replies with a result |
| Where it stops | Hands over to a human | Hands a number to a human | Hands over only what is outside the rules |
| Example | “Your withdrawal is being processed” | “Player #4821: high churn risk” | Finds the stuck payout, opens a ticket in payments, gives the player an ETA |
Six places AI agents already work in iGaming
1. Player support around money
Withdrawals, verification, missing bonuses — the questions where a player waiting too long simply leaves. The agent checks the payment status, opens a ticket in the right department with full context, keeps the player informed and pushes until the ticket is closed, around the clock.
2. VIP service and player retention
The agent notices a player who went quiet, works out the likely reason from their history and writes personally, in the brand's voice. Bonuses go out only within the corridor for that tier; anything above it reaches the Head of VIP with a ready proposal. More on this in our guide to iGaming player churn and on the player care agent page.
3. Payment provider and cashier support
For PSPs serving casinos, merchants ask the same questions about payouts, disputed deposits and report mismatches. The agent answers with a status, a reason and an ETA instead of “we're checking”. See AI in payments: use cases for payment providers.
4. Affiliate reporting and reconciliation
Affiliates live between a tracker, operator dashboards and ad accounts that never agree. The agent collects the data itself, counts the economics of every flow and reconciles operator reports with the tracker row by row. See the affiliate traffic agent.
5. Incident detection from the request flow
When a payment method fails, twenty identical complaints arrive within minutes. The agent groups them into one incident, tells the tech team which method is failing and since when, and warns affected clients with a workaround before the wave of complaints.
6. Back-office routine
Reports, status updates, chasing documents for KYC, routing requests by topic. Each task is small, but together they are what makes support headcount grow with every new market or language.
Where AI doesn't belong (yet)
Being explicit about the limits is what makes an agent safe to run on live players. In our deployments these stay with people:
- Moving money. The agent prepares a payout retry, a bank request or a bonus above the corridor; a manager confirms with one button.
- Responsible gambling decisions. Who must not receive offers and which signals go straight to responsible staff are hard limits in code, not a judgement call for the model.
- Anything it cannot verify. An unreadable receipt is never “read” by guessing digits — the agent asks for the original.
- Unusual cases. Where the agent is unsure, it hands over with context instead of improvising.
Guardrails: rules in code, not in the prompt
A prompt can be talked around; code cannot. Limits, budgets, bonus corridors, contact frequency and bans are enforced outside the model, so a persuasive player can't negotiate past them. Every reply and action is logged and can be pulled up later.
Data is the other half. An agent can run in your own perimeter — cloud, hybrid or fully on your side — with sensitive fields anonymised before they reach the model. For operators that is usually the first question, and it should be answered before any pilot.
How to measure the effect
Run the agent against a control group: part of the requests or players handled by the agent, the rest as before. Compare, not guess:
- reply time and time to resolution;
- share of requests closed without a human;
- cost per case;
- for players — deposits, retention and margin in the agent group versus the control group;
- churn after a problem request (a stuck withdrawal, a missing bonus).
How to start
- Diagnostic. An export of requests and an hour's call show what can be taken off the team and what it costs today.
- Test on past cases. Before launch the agent works through your closed requests, so you see how it would have handled each one and where it would have handed over.
- Pilot on one flow. One request type or one player segment, measured before and after. The scale decision is made on numbers.
Example: an operator's report against the tracker
Reconciles the operator's report with the tracker and prepares the dispute row by row
When the operator sends a payout, the agent checks every sign-up and deposit against the tracker: what the operator didn't count, why, which rows are disputed. The manager gets a ready list with reasons, not a feeling of being underpaid.
Results from our deployments
Live clients under NDA; numbers come from client reports.
Retains players on its own: writes, grants bonuses, solves problems and brings back those who went quiet — in your brand's voice, by your rules, inside your perimeter.