A platform torn down instead of patched
When eConnect decided to build its technology around artificial intelligence, it didn’t add an AI tab to the software it already sold. It rebuilt the platform from the ground up. That single decision tells you more about where on-device AI in gaming is heading than any vendor keynote, because rebuilding is expensive, slow and risky, and nobody does it unless bolting the new thing onto the old thing genuinely doesn’t work.
Henry Valentino, the company’s president and CEO, laid out the reasoning in a video interview with CDC Gaming at G2E. Two parts of his argument matter beyond his own product line. First, AI features grafted onto a system designed for something else tend to stay features, never capabilities. Second, and more interesting for anyone who cares about what casinos know about their customers: the processing stays on the property, behind the operator’s own firewall, instead of being shipped to somebody else’s data centre.
What on-device AI in gaming actually means
On-device AI, sometimes called edge AI casino technology, means the model runs on hardware the operator controls, physically on site or in its own private infrastructure. Video frames, player records and transaction logs are analysed where they’re generated. The answer leaves the building; the raw data doesn’t.
Compare that with the default shape of most modern AI products. You collect something, you post it to an API, a model somewhere else does the thinking, and a result comes back. That design is cheap to build and easy to scale, which is exactly why it dominates. It also means a continuous stream of sensitive material crossing a network boundary into a third party’s custody.
For a cardroom or a slot floor, the material in question is not benign. It’s surveillance footage of identifiable people, loyalty-card histories, buy-ins and cash-outs, employee movements, and increasingly biometric analysis derived from faces. Once that leaves the property, the operator is no longer the only party with a copy.
Why the cloud is the part operators flinch at
Ask a compliance director why they’d pay more for AI without cloud dependency and you’ll get a short list, delivered flatly, because they’ve had the argument before.
- Custody. A casino licence makes the operator answerable for its data. Regulators are unimpressed by “our vendor handles that.”
- Biometrics law. Face-based analysis sits in one of the most litigated corners of privacy regulation. Several US states regulate biometric identifiers specifically, with consent requirements attached, and Illinois’ biometric privacy statute famously allows individuals to sue directly. Fewer copies in fewer places is the cheapest risk reduction available.
- Jurisdiction. Cloud regions are geography. A tribal property, a European operator under GDPR, and a US commercial casino all have different answers to “where may this data physically rest?”
- Breach surface. Gaming has already learned, expensively, what happens when an attacker reaches customer records through a connected system.
- Time. Surveillance decisions are made in seconds. A model that has to make a network round trip before flagging something has already lost the moment.
None of this is anti-cloud ideology. It’s gaming data privacy arithmetic. The question isn’t whether cloud infrastructure is secure, it’s who is holding the bag when it isn’t.
The trade-offs, laid out honestly
On-premises AI is not free, and anyone selling it as a pure win is selling. Here’s the shape of the choice as it actually presents itself to a casino tech infrastructure team:
| Consideration | On-device / on-premises | Cloud processing |
|---|---|---|
| Where raw data sits | Behind the operator’s firewall | In a third-party data centre |
| Response time | Local, no network round trip | Depends on connectivity |
| Cost model | Capital spend on hardware, then fixed | Ongoing per-use or subscription |
| Model updates | Pushed to each site, slower to roll out | Updated centrally, instantly |
| Scaling across properties | Repeat the install per site | Add capacity on demand |
| Regulatory story | Simpler custody and residency answers | Needs contracts, audits, residency controls |
| Failure mode | Site-level outage | Vendor outage hits every property at once |
That middle row is the one vendors quietly worry about. Cloud AI improves every week without anyone noticing. On-premises AI improves when someone schedules the upgrade. Rebuilding a platform for local inference means owning a deployment problem forever, which is precisely why eConnect’s choice to start from scratch rather than retrofit reads as a bet rather than a feature release.
Beyond catching cheats
Surveillance is where casino AI earns its budget, and the pitch writes itself: cameras that already exist, watched by software that doesn’t blink. But the more revealing part of Valentino’s argument is about the data casinos already hold and barely use.
The example he gives is uncarded play. On any floor, a meaningful share of action comes from people who never present a loyalty card, which means the property is effectively blind to who its customers are and what they’re worth. Linking anonymous play to a recognised face or pattern closes that gap. It is also, let’s be clear, the most privacy-sensitive thing on the list, and the reason the on-premises question stops being abstract. Identifying unidentified guests is powerful marketing and a genuine ethical problem at the same time, and the only honest position is that it needs consent frameworks and signage, not just better models.
The less controversial uses are operational: staffing levels against actual floor traffic, machine downtime, queue lengths, incident patterns by time of day. Casinos are drowning in telemetry and have historically turned very little of it into decisions. That’s the real prize, and it doesn’t require anyone’s face.
What it changes for the person at the table
If you play, the practical difference is custody, not visibility. On-device AI doesn’t mean you’re being watched less. It means the recording of you being watched is less likely to be sitting on a server belonging to a company you’ve never heard of, in a country you’ve never visited. That’s a real improvement in responsible data handling in gaming, and it’s also narrower than the marketing will suggest.
Worth knowing: on-premises does not automatically mean well-governed. A badly configured local server is worse than a well-run cloud tenancy. The questions that matter are retention (how long is footage and derived biometric data kept?), purpose limitation (is player identification used for safer-gambling interventions, or only for marketing?), and access (who inside the property can query it?). Those are policy choices, not architecture choices, and no amount of edge hardware answers them.
There’s an upside worth naming. The same pattern-recognition that spots advantage play can flag signs of harm: escalating session lengths, chasing behaviour, deposit patterns that look nothing like the previous six months. Operators already have duty-of-care obligations in most regulated markets, and tools that keep this analysis in-house make it easier to act on without exporting a sensitive health-adjacent profile to a vendor. If you want control over your own data rather than relying on someone else’s, use the levers the operator must give you: deposit and loss limits, session reminders, cool-off and self-exclusion. If gambling has stopped being entertainment, stop and talk to a support service in your jurisdiction.
The test for the next 18 months
Every supplier at the next trade show will claim AI. The useful filter is simple: ask where the inference happens, what leaves the property, how long derived data is retained, and who signs off when the model is wrong. Companies that rebuilt for local processing will answer in one sentence. Companies that added a feature will answer with a diagram.
That’s the actual story in this interview. Not that AI is coming to casinos, which is old news, but that the architecture decision underneath it, cloud or on-device, is turning into a privacy position whether operators meant it that way or not.