SBC Webinar Highlights: Trust, Autonomy And iGaming-Native AI Support

SBC Insights Webinar brought two of our founders together with two operators running our AI platform in production, to answer a question the industry keeps circling: can AI resolve player issues with full autonomy?

The panel featured Alex Jones, COO at Raphie, Alex Gould, CTO at Raphie, Will Colello, VP of Player Operations at ARB Interactive, and Jake Bush, Director of Operations at BetSaracen. ARB Interactive is one of Raphie's largest partners, and BetSaracen is four and a half years into the relationship.

The session settled quickly on a single argument. Autonomy is earned in stages, and how far it goes depends on the data the AI is allowed to reach. Both operators started with a narrow set of low-risk use cases and widened authority over time, with BetSaracen recently granting limited write access to player accounts and reporting roughly 60% of inbound contacts automated today. Raphie positioned 70 to 80% as achievable once integrations mature, with a clear caution that this is a journey rather than a day-one setting.

Responsible gaming marked the one boundary nobody moved: self-exclusion and RG signals trigger an immediate, fully informed handoff to a human agent. Around that sat the case for industry-specific models, from jurisdictional compliance to the data complexity in iGaming. The hour closed on chargeback automation, payments, risk and fraud, with voice flagged as a later horizon.

What the panel discussed

• What true resolution requires in practice: account-level data from the PAM and wagering platform, and the ceiling operators hit with chatbot deflection, running at around 25% for ARB Interactive

• Player support reframed as a retention driver rather than a non-revenue cost line, with deflection treated as a churn risk

• How players and operators tell generic AI apart from genuine resolution: tone, concision, localisation and taking the next concrete action

• Hallucination risk tiers, with P1 cases such as self-exclusion ruled out through multi-model checks and groundedness verification, and lower-priority misreads described as under 1% of conversations

• Graduated autonomy and risk appetite, from simple contact types through to the recent move to limited write access on player accounts

• Why iGaming-specific models matter: responsible gaming safeguards, fine-tuning for bet types such as accumulators, and integration with back office and ticketing systems

• Testing and readiness: automated evals, manual QA staffed by trained human agents, shadow mode, and incremental post-launch rollout

• Monitoring and expectation setting: copilot and operational dashboards, BetSaracen at roughly 60% automation today, and 60 to 80% framed as a journey rather than a day-one outcome

• Where the category goes next: normalisation of AI in daily life, chargeback dispute automation, payments, risk and fraud, reactivation, and eventually voice

Strongest quotes and standout moments

Developing an AI platform with the humans who ran the floor

Alex Jones, COO, Raphie

"Our core business prior to this AI product was providing outsourced operations with human agents. So we are in a fairly unique place, where we can test it with our humans as well."

Alex Gould, CTO, Raphie

"We have our team of human agents from the traditional part of the business, really well skilled and trained in the industry themselves. They have been able to provide a lot of support with what we call manual QA."

Multiple models checking their own homework

Alex Gould, CTO, Raphie

"We insert multiple checks throughout the creation of a response, often using multiple different models to essentially check their own homework. Groundedness checks make sure that every word in the response can be traced back to a factual artifact."

"All of those things, we work really hard to rule out essentially any P1 hallucination."

What great AI sounds like

Alex Gould, CTO, Raphie

"Badly implemented AI just dumps a whole lot of information on the player. Really well implemented AI is 'less is more', being able to condense the response down to the very thing the player is looking for."

"Some of our best chat sessions are when players start speaking to the AI like it's a human, saying things like, hope you have a good day."

Alex Jones, COO, Raphie

"The key part of it is providing that human-like experience. We didn't want operators to be able to decipher between talking to one of our human agents or the automation agent."

Support is a player retention engine

Alex Jones, COO, Raphie

"Customer support is quickly seen as a non-revenue-generating vertical for an operator. Take a step back and good customer support helps you retain a customer. It's a lot easier to retain a customer than to acquire one."

"Deflecting a customer over and over again is just going to provide a poor experience, which leads to churn."

Why generic models can break on a parlay bet

Alex Gould, CTO, Raphie

"To a player familiar with them, an accumulator is quite simple. On the data side, the amount of data associated with a bet of that type is large and quite confusing. Generic AI doesn't handle that kind of bet well, and we saw lots and lots of errors using something out of the box."

Alex Jones, COO, Raphie

"They advertise these high resolution numbers, but you can't get those unless you have deep integration with your wagering technology to get that account-specific data."

Nothing should feel like a leap of faith

Alex Gould, CTO, Raphie

"None of this should feel like taking a huge leap of faith, and nothing should be an overhaul. Find the thing we can move fastest on, with the most controlled or simplest use case, and be incremental from there."

"Everything in AI is changing so fast now that you need to shorten the time frames you're thinking about. Break it up into smaller pieces."

Automation is a journey, not a launch setting

Alex Jones, COO, Raphie

"Operators need to understand that they shouldn't expect to get 60 to 80% automation from day one. It happens over time, understanding how your players respond to the agent and fine-tuning from there."

"What are you comfortable with, and what amount of data can you provide to the agent? That's where you're going to get that true resolution."

Full session summary

1. Introductions and session framing

Charlie Horner opens the SBC Insights session and sets the central question: AI is transforming player support, and the issue is whether operators can trust it to handle complex interactions with no one overseeing it. Each participant introduces themselves. Will Colello notes ARB Interactive is currently Raphie's largest partner with hands-on experience of the tools. Jake Bush notes roughly four and a half years working with the team.

2. What autonomous resolution looks like in practice

Account-specific data as the precondition. Alex Jones argues true resolution requires pulling account-level information from the PAM or wagering platform, covering withdrawals, deposits and wagering queries. Deflection-style FAQ responses fall short of this.

Why iGaming queries are more complex. Alex Gould contrasts retail-style queries, such as fabric or sizing questions, with iGaming, where resolving an issue almost always requires player context. Deep integration and a clear map of the player journey become essential.

Moving past chatbot deflection. Will Colello describes deflecting around 25% of contacts through an internal chatbot, and positions the AI agent as the next layer: accessing the back office, applying real-time information, and resolving first time. He links this to staffing levels, speed of experience, and safe words that let the operator step through autonomy at its own comfort and risk level.

3. The status quo challenge for operators

Support as a cost center. Alex Jones describes customer support being treated as a non-revenue-generating vertical, and reframes it as a retention mechanism. Retaining a player is cheaper than acquiring one, which matters more in mature, share-of-market jurisdictions.

Cheaper delivery while keeping quality. Automation lowers cost while still producing responses that resolve the query, since deflection creates a poor experience and drives churn.

4. Distinguishing generic AI from genuine resolution

Tone and natural language. Alex Gould identifies the familiar tells of poorly implemented AI, including double hyphens and overly verbose writing, and describes the work needed to counter them: reference transcripts, style and tone guides, and heavy QA. The marker of success is players speaking to the AI as they would a person.

Less is more. Badly implemented AI dumps lists and full articles at the player. Well implemented AI condenses the response down to the specific thing the player asked for.

Operator view: localisation. Will Colello notes the Southern US voice of their player base, and that players feel they are speaking to someone in Arkansas.

Operator view: taking the next action. Jake Bush uses pending redemptions as the example. Basic AI explains why a redemption is pending. Raphie identifies the missing document, explains where to upload it in the platform, and provides the link, which is what acting as a legitimate agent means.

Human-like experience as a design principle. Alex Jones explains this was a founding product decision: operators should be unable to tell a human agent response from an automation agent response, which builds trust with the client base and wins new clients.

5. Hallucination, risk tiers and acceptable error

Operator perspective on risk levels. Jake Bush separates an RG or self-exclusion concern from a typical player request, and confirms that RG cases trigger an immediate, fully informed handoff so the human agent knows exactly what they are picking up. He reports the platform has worked flawlessly in his experience.

How the platform mirrors those risk tiers. Alex Gould describes P1 hallucinations, such as anything touching self-exclusion, being ruled out through multiple checks across the response lifecycle, different models checking each other's work, and groundedness checks tracing every word back to a factual artifact.

Lower-priority errors and their frequency. Further down the risk ladder, the AI may misread what a player is asking and answer differently. Alex Gould puts this at under 1% of the time, and cautions against assuming a single observed error is affecting every player.

Safeguards and escalation speed. Will Colello emphasises that the platform's job in sensitive cases is to hand off, and praises the instant handoff from AI to human agent.

6. Disclosure and how trust actually forms

Operator choice on disclosure. Alex Jones notes it is the operator's decision whether to tell players they are speaking to an AI agent.

Audience question from Carla. Carla asks whether trust is eroded by not revealing the AI to the player. Alex Gould responds that in almost all implementations the handoff to the Raphie agent is stated explicitly. The observed pattern is that players relax quickly even when they know, and respond in a human-like way when the AI interacts empathetically, which builds trust without constructing a fake human.

7. The role of humans and graduated autonomy

Use-case-by-use-case onboarding. Alex Jones describes a white box service: agreeing a set of use cases the operator is comfortable with, proving value there, then expanding as the partnership develops. He cites around 60% of contacts automated for certain use cases across Will and Jake's operations, with harder cases still going to humans.

Risk appetite and write access. Jake Bush distinguishes simple high-frequency contacts such as password resets from write ability that makes changes in the back of the account. BetSaracen took an incremental approach, gradually widening responsibility, and has recently granted limited write ability. Anything RG, self-exclusion or otherwise sensitive is escalated to a human by internal policy.

The T1 agent framing. Will Colello treats the AI as a T1 agent covering redemptions, passwords and account access, which frees human agents for the work the operator wants them spending time on and prevents them being inundated.

Data access as the ceiling on automation. Alex Jones ties automation potential of 70 to 80% directly to back-end, player account system and ticketing access.

8. Why iGaming-specific models matter

Risk and safety. Alex Gould places this first: self-exclusion has no analogue in other industries, and jurisdictional compliance and RG standards require safeguards and fine-tuning that generic AI does not provide.

Fine-tuning for industry-specific situations. The accumulator, parlay or multi-bet is the worked example. Simple for a player, and large and confusing on the data side, with varying descriptions and locations. Out-of-the-box models produced many errors on this bet type.

Integrations as the deciding factor. Described as the most important element: connecting to back office and ticketing systems unique to the industry, without which the player cannot be satisfied.

Operator perspective: tracing game errors. Will Colello adds the casino-side equivalent, such as a game crashing mid bonus round, which needs a specific toolset and knowledge of where to look.

Operator perspective: hallucination reduction. Jake Bush argues models built for casino, sports and iGaming lower hallucination risk and answer narrow, specific questions that generic models would likely get wrong.

Generic vendors and headline resolution rates. Alex Jones references operators trying generic helpdesk and CRM solutions that advertise high resolution numbers, and argues those numbers are unreachable without deep wagering platform integration.

9. Testing, QA and readiness for live deployment

Automated evaluations. Alex Gould describes evals, using AI to test the AI at scale, as one of two main testing branches before anything goes live.

Manual QA as a structural advantage. The second branch is traditional manual testing, staffed by Raphie's team of trained human agents from the legacy outsourcing business, reviewing transcripts and AI output.

Shadow mode. Alex Jones explains the platform can run against real-time incoming contacts without sending responses, capturing the data so the operator can see how it would have performed before going live.

Incremental post-launch rollout. The third phase sits after go-live: starting in copilot mode or with one small use case, then stacking use cases as confidence grows.

Operator monitoring in production. Will Colello points to reporting that tracks effect during the early trust-building lifecycle, plus the copilot loop where human agents feed reliable responses back into the agent. Jake Bush describes watching the product through the full copilot dashboard phase, where responses were graded and improved, through to the operational dashboard, and confirms roughly 60% of use cases automated today.

10. Round table: top advice before granting autonomy

Alex Jones. Start from the operator's comfort level, and understand upfront what data can be given to the agent, since data access determines whether true resolution is possible.

Will Colello. Arrive with a strategy. His own was weeding out repetitive top contact types so human agents could focus on complex inquiries.

Jake Bush. Define the authority you are willing to give it, start with a few use cases, and expand as comfort level and risk appetite allow.

Alex Gould. Be incremental, for two reasons: nothing should feel like a leap of faith or a full overhaul, and AI is moving fast enough that planning horizons should shorten.

Expectation setting. Alex Jones adds that 60 to 80% automation develops over time as player interaction data accumulates and use cases are added.

11. Audience question: detecting nuanced responsible gaming signals

The question. Gabriella, working in responsible gaming, asks how AI systems detect RG situations when players rarely use explicit terms such as addiction or self-harm.

Curated examples over keyword matching. Alex Gould describes curating examples relevant to the operator's business and player type, much as you would brief a human agent on what to look for and the nuance of each situation, then using those to fine-tune the models. Keywords alone are too blunt. The example given is "lost everything" followed by laughter versus "lost everything" as a flat statement.

Operator example. Jake Bush describes compliance team interest in this, and a real pattern where a player pushing for a faster withdrawal mentions needing to pay a bill, which the system escalates so a human agent can read the temperature of the situation.

Why phased release supports this. Will Colello links accuracy here to training investment, processing past conversations, and releasing in waves with concentration on specific aspects first.

12. The future of trust and where the category goes next

Normalisation of AI. Alex Jones observes that hesitation about speaking to an AI agent was more common 6 to 12 months ago, and daily use of consumer models is making it normal, which reduces operator fear of pushing players away.

Player trust trajectory. Will Colello foresees most or all tier one inquiries being handled by AI, drawing a parallel with how readily people now run everyday questions through consumer AI tools and accept the answers.

Expansion into payments, risk and fraud. Jake Bush points to chargebacks, where a single case can take 30 minutes to an hour, as an obvious target for deeper integration with a final human review before bank submission. Alex Jones confirms a chargeback dispute automation product already exists, alongside ambitions across reactivation, payments, risk and fraud, covering the full player lifecycle.

Voice. Jake Bush flags voice as a later opportunity once languages and dialects are tuned.

Honesty and updated assumptions. Alex Gould closes on continuously reassessing preconceptions, since what hallucinated constantly a year ago has since been fixed, and encouraging clients to bring data and examples rather than relying on old impressions.

Partnership framing. Alex Jones frames the relationship as partnership: bring the problem, resolve it together, expand use cases from there.