Payment Agents: Your payment team, always on

Meet the team

AI agents built into Tonder to continuously optimize how your payments perform. But the job titles aren’t a metaphor — they do what those roles actually do at companies large enough to staff them: the daily performance review, the decline investigations, the fraud case files, the “should we change this rule?” analysis.

That last part is the whole point. A payments analyst and a risk analyst are two hires most operators in this vertical cannot justify — not because the work isn’t worth doing, but because it only pays for itself once you’re large enough to have already lost the money it would have saved. So the work doesn’t get done. Approval drifts for a month before anyone notices. A card tester runs for three weeks because nobody was looking at velocity on a Tuesday.

Most providers answer this with a dashboard. Dashboards are where questions go to die. You log in when something feels wrong, you stare at a chart, and you leave with a vague sense that approval “dipped a little.” Nobody opens a dashboard at 6 AM. Everybody reads Slack.

Marcus posts before you wake up

Payment problems don’t announce themselves. Approval doesn’t fall over; it drifts, two points at a time, until a month has gone. A card tester doesn’t trip an alarm; it runs for three weeks and then becomes chargebacks. An issuer doesn’t tell you it changed its overnight risk posture — it just starts saying no after 10 PM. By the time any of this is obvious enough to investigate, the money is already gone.

Which is why the important thing about Marcus isn’t that he can answer questions. It’s that he goes first. The briefing lands at six in the morning whether or not anyone thought to ask, whether or not anything looked wrong yesterday — and the numbers are in the message, not behind a link.

Everything Marcus watches, in one list:

Where the numbers come from

A fair question about any agent that writes confident sentences: is it computing this, or narrating it? Marcus is not summarising a dashboard export. He reads the authorization record — the attempt level, not the settled level — and compares it against your own history, not an industry average.

The report is the beginning of the conversation

Here is the part that changes how teams actually work: you can talk back. Marcus isn’t a broadcast bot. Reply in the thread the way you’d ask an analyst sitting next to you — plain language, no syntax, no query builder — and the answer comes back with the data behind it, in seconds rather than in next week’s meeting.

Look at what the second question actually was. That’s not reporting — that’s segment analysis that used to mean filing a ticket with a BI team and waiting a week. It happened inline, in the thread, while the question was still hot. And because both agents read the same data, Marcus can hand a question to Zeno mid-conversation: an approval anomaly becomes a fraud investigation becomes a rule change, without anyone exporting a CSV.

Three things that turned up this way

All three come from the same demo operator used throughout this article — not customer results, and each one is checkable against the reports above.

Zeno: the memory your team never had

A blocklist that ends at your own front door is a blocklist a fraudster only has to beat once. They burn you on Monday, walk next door on Tuesday, and arrive as a new customer with a clean history — because as far as that operator can see, they are one. Every risk team in this market has been fighting the same people independently, and paying to learn the same lesson separately.

Marcus tells you what happened. Zeno tells you who you’re dealing with — and his memory doesn’t stop at your front door.

Think of it as a credit bureau for players. Every deposit, decline, chargeback and device signal builds a profile, and because Tonder processes for multiple operators in the same vertical, that profile survives the move from one brand to the next.

Fraud cases, not event logs

When something like the card-testing case fires, Zeno writes the case file: every card attempted, the BINs and issuers behind them, device and velocity signals, the timeline, and whether that fingerprint has been seen at other operators on the network. Your team gets a story with a recommendation at the end, not a spreadsheet of raw events to interpret.

A blocklist that remembers why

Confirmed bad actors stay blocked across email, card fingerprint and device. The difference from a homemade list in a spreadsheet is that every entry carries its reason and its history — so six months later nobody is asking “why is this one blocked?” and quietly unblocking a chargeback machine.

And a whitelist, because false positives cost more

Risk teams love to talk about the fraud they stopped and never about the VIP they annoyed. A player with 40 clean deposits and zero chargebacks shouldn’t meet the same friction as an account created twenty minutes ago. The whitelist means your proven players get the frictionless path — fewer challenges, fewer manual reviews, fewer abandoned deposits from exactly the people who fund the business.

Underneath all three sits a score that resolves into green, yellow or red, and that score drives routing automatically — friction goes where the risk is and nowhere else. The scoring engine is a subject of its own; we wrote it up in Guardian AI for iGaming.

What a Zeno report looks like

Monthly, per property, to whoever owns risk. The first page is the one most people read:

Two things to notice. The findings are ranked by recoverable pesos, not by severity theatre — the compliance breach comes first because it has a deadline attached, and the 40.9% issuer comes second because it is the largest pile of money on the floor. And the numbers that improved are stated as plainly as the ones that didn’t; a fraud report that only ever brings bad news stops being read by March.

Then the same thing happens as with Marcus: you reply to it. “Which disputes are still in flight?” “What would EFM compliance cost us in blocked volume?” The report is a starting position, not a verdict.