For licensed online casino operators
Agents run the floor. You make the calls.
AI Pit Boss is an operations cockpit. AI agents handle first-line payments, compliance, support, VIP, CRM and games work inside guardrails. Your people make the calls that money, the law and your licence reserve for them, from one control room.
Your table · 2 decisions
Decision cards are illustrative examples of the pack format.
Agents, last 24 hours
- Host calls booked1,091
- Re-KYC forced718
- Player messages sent693
- Win-back offers459
- ID requested388
- RG limits applied273
The problem
Alerts, not decisions. Desks that don’t join up.
open alerts
In our sandbox, about 2,500 were marked critical, because one aged withdrawal raised one critical alert. Nobody can work a list that size. Sandbox · synthetic players · measured 30/10/2026.
agent actions to person actions
Over seven days the agents took about 4,800 actions and a person took one, because the sandbox has no staff. The queue only grows when decisions have nowhere to go. Sandbox · synthetic players.
coverage is a floor
Keeping one person on every shift of a 24/7 desk takes about 5.2 full-time staff, whatever the workload. Modelled, assumptions (leave, sickness, 37.5 h weeks).
Regulators are fining joining-up failures
Three cases from the public record. We cite what the regulators and law firms published, and note where a case is alleged.
QuinnBet · UKGC
£609,104 payment in lieu (20/08/2026), including £193,118 disgorgement. Harm reviews were occasionally delayed a second day because of staffing, loss-limit alerts fired only on a further deposit, and a platform migration broke deposit limits for 194 customers.
CMS enforcement round-up, autumn 2026 (law-firm summary)
theScore · AGCO
C$105,000 in penalties, announced 07/10/2025. The notice says a patron showed signs of distress to the operator’s VIP host, and that this did not reach responsible-gambling monitoring. These are allegations and can be appealed.
ProgressPlay · UKGC
£1m (reported 08/2025) for anti-money-laundering failures, including not checking source of funds. A repeat case: a 2022 penalty was £175,718.
CasinoBeats report (trade press)
In these cases the failures differed: staffing, hand-offs between desks, detection design and missed checks. We do not claim a cockpit would have prevented them.
How it works
Agents deal. You decide.
- Agents deal the routine workStatus replies, ticket triage, document requests and payments under the limit with every gate passed run on their own, and every action lands in a ledger.
- You get a decision, not an alertAlerts roll up inside the agents. People see grouped decisions, sorted by clock then money at risk, with a budget per person per hour.
- The decision packEvery decision shows the same pack: the proposed action, why (the rule plus the one piece of evidence that matters), what the agent is unsure of, what happens on reject, and whether it can be undone.
- Explain, with the rule quotedAsk “what does the UK require here?” and get the facts on file plus the jurisdiction’s rule quoted with its reference and date. Explain makes no recommendation on decisions reserved to Compliance and refuses questions outside its rulebook.
- One player timelineResponsible gambling, AML, KYC, support, VIP notes, CRM sends and payments on one timeline, with role-scoped views. One risk state (clear, watch, suppress, excluded) gates offers, marketing and autopilot everywhere.
- Audit you can exportReason codes on every human decision, shift handovers with sign-off, and a tamper-evident ledger exportable against hashes.
The desks
A casino floor your whole operation can read
Each desk is a felt table with the plain name beside it. Colour means exception: red is act now, an amber ring is watch, no colour is normal.
Payments
The Cage
Withdrawal worklist with jurisdiction rule badges, a decision pack for each item, amount-gated second reviewer, PSP health.
AML · responsible gambling
Compliance
Case worklist with clocks, evidence side by side, harm markers in minutes not next morning, SAR drafted by agents and filed only by your MLRO.
Support
Guest services
Agents answer read-only status questions and draft replies. Team leads see SLA breaches and contacts flagged as harm or AML, with tipping-off safeguards.
VIP
Host desk
Today’s call list, a pre-offer compliance gate, and host notes that flow into the player’s responsible-gambling timeline.
Marketing · CRM
Promotions
A pre-launch audience gate that removes excluded and harm-flagged players, and a readout of bonus cost against a holdout.
Games
Games floor
Game and provider monitoring, with requirements still being specified. Planned for a later release.
Safety and compliance
What agents may do alone, and what always needs a person
| Action class | Agent alone | Needs a person |
|---|---|---|
| Apply a self-exclusion, block a register match, restrict on a strong harm marker | Always, immediately (safety floor) | Review afterwards |
| Lift a self-exclusion or make an exception | Never | RG lead plus a second person |
| Apply an RG limit or cool-off on a policy trigger | Yes | Per-customer review; the player can contest |
| Pass affordability or source-of-funds checks | Never | KYC/AML analyst |
| Dismiss an AML alert as a false positive | Never | AML analyst, sampled by the MLRO |
| File a SAR or close an account on AML grounds | Never (drafts only) | MLRO |
| Pay a withdrawal | Below the limit, all gates passed, no risk state | Above the limit or with any trigger |
| Offer a bonus or comp; grant VIP status | Never to a watch, suppress or excluded player. VIP status never | Host or CRM lead within policy |
| Status replies, ticket triage, document requests | Yes | Anything touching harm, an AML hold or money |
Fleet pause with a safety floor
Anyone can pause; only a named owner resumes. New actions stop within 30 seconds, and actions in flight finish or roll back. Self-exclusion and harm restriction keep running and fail closed.
Tamper-evident audit
Every autonomous action and human decision is logged with a reason code. The ledger exports and verifies against hashes. Records of Advise exchanges are kept for five years.
Human review and contest
Every agent action that affects a player gets a per-customer review, and the player has a route to contest it (GDPR Art. 22 and the UK equivalents).
Jurisdiction rulebook
A versioned store of rule texts per jurisdiction, with effective dates, source links and two-person sign-off. Explain answers only from the rulebook plus your data. Stale rules are flagged.
These describe the product design. What is built today and what is specified is confirmed in pilot scoping. Read the security page.
Proof from the sandbox
What we have measured, and what we have only modelled
We have no paying operators and no case studies. The evidence is a sandbox of synthetic players and a staffing model. Both are labelled.
agent actions a day
4,623 applied, 109 rejected, 38 failed and 28 pending in the 24 hours to 30/10/2026, across three brands at 0.4× volume.
operations FTE
For a reference operator at €5M gross gaming revenue a month. Range: 100–131 today, 66–86 with the cockpit. Arrivals are measured in the sandbox; handling times and automation shares are our assumptions.
FTE saved, all-assumptions-together
If every assumption lands low, middle or high together, the saving is 10, 32 or 94 FTE. That spread is the honest uncertainty until operator interviews replace our ranges.
The model also says some desks grow
Per-customer review of agent responsible-gambling actions (about 680 a day in the model) takes that desk from roughly 9 to roughly 14 FTE. Agent autonomy creates review work, and the product has to make that review fast and honest. Payments falls from about 20 to about 5 FTE, which is mostly a 24/7 coverage floor.
The pilot
Eight weeks in shadow mode
Agents propose. Your people decide as they do today, in their own tools. We measure how often they agree, per action class. Then you switch agents on, desk by desk, only where the agreement supports it.
Pricing
Fixed monthly pricing. No revenue share.
We price the cockpit, not your players’ losses. Talk to us for a quote scoped to your desks and markets.
FAQ
Questions operators ask
What data do you hold?
In the sandbox, only synthetic players. In a pilot, decisions write through to your platform, which stays the system of record. The cockpit does not keep its own copy of money state. See Security for the data-handling principles.
How does it connect to our systems?
Through your PAM or back office, via an executor that carries the rate limits, a kill switch and an audit log. If the cockpit or a model provider goes down, your back office keeps working, agents stop at the safety floor, and a runbook says who does what. Integration scope is agreed in the pilot.
Which markets does it cover?
The design carries a versioned rulebook per jurisdiction. The research behind it covers the UK, Ontario (Canada), Sweden, Brazil, Malta and Curaçao. Which rules are loaded for you is confirmed in pilot scoping, not assumed.
Which model providers do you use?
We do not publish a provider list yet because we have no production deployment. The provider, hosting region and data-processing terms are set with each operator before a pilot starts, and every agent exchange is recorded.
What stays human?
Filing a SAR, MLRO sign-off, closing an account on AML grounds, lifting a self-exclusion or making an exception, dismissing an AML alert, passing affordability or source-of-funds checks, and granting VIP status. Agents never do these alone. See the matrix above.
Is it secure?
Role-scoped views, a tamper-evident audit ledger, a DPIA before any release to an operator, and a need-to-know compartment for MLRO data. We have no security certifications to claim. See Security for what is built into the design and what is still to be confirmed.
Do you replace our compliance team?
No. The cockpit is not compliance advice. It takes first-line work and brings decisions to the people accountable for them.