The CABI framework
Conversational Augmented BI, built for regulated operators.
Your teams ask in plain English. The model proposes; deterministic checks decide what runs. Answers come from your governed data layer, with the evidence attached.
The structural problem
Why fast questions still get slow answers.
Operators rarely lack data or dashboards. What they lack is a short, governed path from a question to an answer finance will sign off.
Copy sprawl
Every tool takes its own copy of player data. Each copy adds an ETL job, a refresh delay and another system to cover in a security review.
Definition drift
Finance, CRM and the board pack each compute NGR their own way. Meetings start by reconciling the number instead of acting on it.
The request queue
Ad hoc questions wait behind a small BI team. A VIP drift signal that arrives ten days late has lost half its intervention window.
The framework
Four layers, one governed path.
Each layer has one job, and they only work together.
- 01ConversationAsk in plain English. Each question is routed as a number, a definition or a why, and Bounty asks for the period or scope when it is missing.
- 02AugmentationAnswers arrive with explanation: investigations that test each possible driver, narrated reports and anomalies surfaced from live metrics.
- 03GovernanceSemantic contracts define what the model may query. Deterministic checks validate every generated query before it runs, read-only.
- 04ActionInsight becomes a ranked decision: one Next Best Action per player, Responsible Gaming cases with an audit trail, campaigns with fatigue guards.
Lean by design
Domain-adapted intelligence, not a giant general model.
Because governance does the heavy lifting, the language work runs on compact models that know your industry. Answers stay fluent in the business, and running costs stay small as more people ask.
Fluent in iGaming
Compact models adapted to the industry, grounded in your governed definitions, so FTD, NGR and a drifting VIP mean exactly what your team means by them.
Right-sized by design
Deterministic code carries the definitions, checks and charts, so the model only does the language work. Larger models are kept for the few steps that need them.
Cost that stays small
Around 1,500 business questions answered in four months, at an inference cost that is a rounding error next to the analyst time they replace.
Mechanics, not magic
How a question becomes a governed answer.
- 01Route the questionA number goes to governed query generation, a definition to the metric registry, a why question to an investigation.
- 02Load the contractsEvery queryable table is described in a versioned contract. Stale copies are listed as forbidden.
- 03Validate the queryEvery table and field is checked against the contracts. One repair attempt, then a query that still reaches outside them is rejected, never run.
- 04Run it read-onlyOnly read queries are admitted, with bounds on time and size.
- 05Draw from the rowsThe chart comes straight from the returned rows. The model writes the summary, and the table is one click away.
Team by team
What changes on Monday morning.
VIP and CRM
Scoring nightly or as players act, ranked action buckets and one Next Best Action per player, so drift reaches the relationship manager before Day 30.
Finance
GGR, NGR, FTD and bonus cost defined once, with fee bases and thresholds written down, so a difference traces to a stated assumption.
Compliance
RG cases with a full audit trail, a KYC verification funnel, a bonus-abuse review queue and an audit log of access changes.
Executive and BI
Follow-up questions answered in the meeting on board-pack definitions, while the BI team maintains the contracts self-service runs on.
The white papers
Read the framework, then the architecture.
Two papers: one for the business case, one for the people who will have to defend the architecture.
White paper · 18 pages
Conversational Augmented BI: closing the gap between the question and the governed answer
The business case for CEOs, CDOs, VIP and CRM Directors and Compliance Officers: the structural problem, the four-layer framework, what changes for each team and how to run a pilot.
Technical white paper · 23 pages
CABI reference architecture: governed text-to-SQL on a semantic layer
For CTOs and Heads of BI: semantic contracts, the SQL validation gate, execution safety, the investigation engine, scoring, access control and the known limits of each.
FAQ
CABI questions
Does it get expensive as more people use it?
No. The governed layers do the heavy lifting, so the language work runs on compact models adapted to iGaming. In production, around 1,500 business questions over four months cost a rounding error in inference next to the analyst time they replaced, so you can open conversational access to every team.
What is CABI?
CABI (Conversational Augmented Business Intelligence) is a framework for governed, conversational analytics, first drafted at Bounty AI in 2024. It has four layers: conversation, augmentation, governance and action. Bounty AI is its working implementation for regulated iGaming operators.
How is CABI different from putting a chatbot on the warehouse?
A model writing SQL from raw schema guesses join paths, invents columns and reads stale tables. In CABI the model only proposes: semantic contracts define what exists, a validator checks every query against them, and only read queries run. The chart is drawn from the returned rows, not generated.
Can the written answer be wrong?
Yes. The summary is written by a language model, which is why every answer shows the chart and the table it came from, the period and the row count, with a standing note that AI can make mistakes. A stricter check that verifies every figure in the summary against the data is built and on the roadmap.
Does CABI forecast?
Not today. Forecasting and what-if analysis are on the roadmap and are not part of the current product. The white papers mark every roadmap item clearly.
Which data sources does it connect to?
All of them. Platform and payment databases, an existing warehouse, exports, files and APIs are mapped into one governed data layer, and questions run on that layer rather than on an exported copy per tool. Ask us for the exact data path you would deploy.
Is Bounty AI certified by a gaming regulator?
No. Bounty AI does not claim certification or approval by NJ DGE, PA PGCB, UKGC, MGA or any other regulator. It supports compliance teams with current data, review queues and audit trails; regulatory obligations stay with the operator.
Keep reading
The questions buyers ask next.
See CABI on your own data.
A pilot covers two or three workflows you choose, with GGR, NGR and deposits reconciled against your finance figures.
Pilot: your data, in your environment. Demo: sample data, access reviewed within one business day.