How much autonomy should AI have in a casino back office?

“Is the AI on or off?” is the wrong question for a casino back office. The useful question is what the AI is allowed to do, with what data, and who approves it. We use five levels.

The five levels

1. Level 0, observe. The AI reads and explains. It changes nothing.
2. Level 1, recommend. It proposes an action and the reasons for it.
3. Level 2, prepare. It prepares the change and shows a preview: audience, cost, sample, wording.
4. Level 3, execute with approval. It asks an authorised operator and performs the action once approved.
5. Level 4, autonomous. It acts alone, only for explicitly permitted, low-risk actions inside set limits.

Levels are granted per capability and per role, not as a global switch.

The line that level 4 never crosses

Whatever the size of the action, these stay at level 3 at most:

• moving money to a player;
• changing a player's ability to play;
• sending a message to a live player;
• touching responsible-gambling limits.

These are the actions where a mistake reaches a person, a regulator or both. Volume does not change that: ten small mistakes are ten reaches.

Explain before you execute

An answer that only repeats a dashboard number is not the AI's work. The format is: the problem, the evidence, the impact, the recommendation, the expected result, and the action with a preview and explicit approval. “Card approvals on one route fell” becomes “approvals fell, these attempts were affected, this is the estimated loss, send card deposits through the other route until it recovers.”

Measure every intervention

Each action is meant to be recorded as before, intervention, after, so that later you can answer whether it worked, what it returned, whether conversion and retention improved, whether risk grew, and whether the recommendation was a mistake. Without this, autonomy cannot be raised responsibly.

Keep the rules out of the prompt

Wallet, ledger, wagering, eligibility, risk and limits are deterministic services. The AI calls them through audited tools. If a rule lived in a prompt, nobody could prove what the system would do tomorrow. See how this differs from a classic back office, or the platform.