Live Chat Prompts Raise Average Bet 14% in 800 Tables

Across a 90-day trial spanning 800 live-dealer tables, tables that received automated, context-aware chat prompts recorded a 14.2% higher average bet per round than a matched control group, according to data shared by two platform vendors and one independent auditor. The effect held across roulette, baccarat, and blackjack variants, though it decayed by roughly a third within six weeks of a table going live. The finding is narrow, but it sharpens a question the live-casino sector has largely answered by assertion rather than measurement: does the chat box move money, and if so, whose money, and for how long?

What the trial actually measured

The 800 tables were split 400/400 between treatment and control, matched on stake distribution, dealer shift, table limit, and timezone. Treatment tables ran a prompt engine that injected a short dealer-side message at intervals tied to round completion rather than wall-clock time — a design choice the researchers say was made to avoid prompts landing mid-decision, which an earlier pilot had shown to irritate players and depress session length.

Average bet per round, not total handle, was the primary endpoint. That distinction matters. Handle rises trivially if session length rises; the trial's headline figure is a per-round measure, which means the lift came from players choosing larger chips or pressing bets, not from staying longer. Secondary endpoints told a more mixed story: session length rose 4.1%, while the share of players who reduced their stake after a prompt (a "cooling" response) sat at 6.8% in treatment versus 5.9% in control.

The numerical anchor worth holding onto: the lift was 14.2% on average, but 21.6% in the first 14 days of a table's life and 9.4% in days 43–90. Whatever the prompt engine is doing, it is not doing it consistently.

The prompt taxonomy

Not all prompts behaved alike. The auditor grouped them into four families:

• Social/neutral ("Good luck, Mr. Okafor — that's three reds in a row"): +9.1% average bet, stable across the 90 days.
• Congratulatory (celebrating a win): +18.7% immediately, decaying fastest, with the largest cooling share at 11.2%.
• Quasi-informational ("The table's running hot tonight"): +16.3%, and the family most likely to draw regulatory attention in jurisdictions that treat dealer statements as advice.
• Return/retention ("Good to see you back"): +11.4% but concentrated almost entirely in the first three sessions after a lapse.

The congratulatory family is where the ethical friction sits. A dealer script that amplifies a win state is, functionally, a nudge toward the cognitive distortion known as the hot-hand effect — the belief that a run of wins predicts further wins. The trial does not claim causation on that mechanism; it measured bet size, not belief. But the decay pattern is consistent with it: the players most affected are the ones whose judgment is most recently anchored to a favourable outcome.

Why the effect decays

Three explanations have been offered for the six-week bleed-off, and they are not mutually exclusive.

Habituation. The prompt engine's phrasing is finite. Players who see the same congratulatory template four times in an hour stop registering it as social contact and start registering it as furniture. The trial's own logs support this: unique-prompt ratio in week one was 0.81; by week six it had fallen to 0.34.

Selection effects. Tables that opened during the trial attracted a mix of new and returning players. As the cohort aged, the composition shifted toward regulars, who are more stake-stable and less responsive to novelty. The control group showed the same drift, just smaller — a reminder that live tables have their own natural maturation curve.

Dealer fatigue. Dealer-side adherence to the prompt schedule fell from 94% in week one to 71% by week nine, per the auditor's spot checks. Dealers reported that scripted prompts felt "false" once they recognised returning players. This is not a trivial operational footnote: it suggests the measured effect is partly an artifact of novelty on the dealer side as well as the player side.

The 800-table number, examined

A sample of 800 tables sounds robust until you ask how many were actually live for the full 90 days. The answer is 612. The remaining 188 opened late or were retired early, and the vendors' headline figure is computed on the full set with imputation for missing days. Recomputing on the 612 fully-observed tables drops the lift to 12.8%. Still positive, still statistically distinguishable from zero at conventional thresholds, but a reminder that "800 tables" is a marketing-friendly denominator.

What the numbers do not settle

Three gaps matter for anyone reading this as operational guidance rather than as a curiosity.

First, the counterfactual is weak. The control group received no prompts at all. A more informative design would compare scripted prompts against unscripted dealer chat, which is what most live tables already have. If organic dealer conversation produces a similar lift, the finding is about interaction, not about automation — and the product case for prompt engines weakens considerably.

Second, the harm side is under-measured. Average bet per round is a revenue metric. The trial reports cooling responses and session length, but not loss-chasing behaviour, deposit frequency, or the share of players who hit a self-imposed limit. A 14.2% lift in average bet is neutral in isolation; paired with a rise in limit-hitting, it would read very differently. The auditor flagged this as the study's largest limitation.

Third, jurisdictional variance is unaddressed. Quasi-informational prompts of the "running hot" type are already restricted in several regulated markets, and the trial's own data show they are among the more effective families. Operators in those markets are looking at a smaller addressable lift than the headline suggests.

A note on where this sits

Live dealer is the fastest-growing vertical in several regulated markets precisely because it feels less automated than RNG tables. That feeling is the product. Prompt engines monetise the feeling directly, which is why the ethical question is not whether a 14.2% lift is real — the trial's design, with its caveats, supports that — but whether an operator should engineer social contact whose measured purpose is to increase stake size. Responsible-gambling frameworks built around deposit limits and time-outs do not currently capture this category of influence, because it operates below the threshold of what those tools measure.

The question the trial leaves open

If the lift decays to roughly 9% by day 43 and dealer adherence is falling faster than player responsiveness, the rational operator's move is not to scale prompt engines but to rotate them — new tables, new scripts, new cohorts — which converts a retention tool into a churn-and-acquire tool with a shorter half-life and a higher regulatory profile. Whether that rotation is sustainable, or whether regulators will treat scripted dealer prompts as a form of inducement requiring disclosure, is the next measurement problem. The 800-table dataset answers a narrow question well. It does not answer whether the sector wants the answer it produced.