When Hub2 came to us, they had two distinct operational problems that shared the same root cause: too much volume, not enough automation. Hub2 runs fintech infrastructure for merchants across Ivory Coast, and their support team was processing every single ticket manually through Zendesk. Each one required the same sequence of steps regardless of complexity: look up the account, check the transaction, compose the response, update the fields, close the ticket. At 2,000+ tickets a month, this consumed the team entirely. There was no triage, no prioritisation, no way to separate a simple status check from a genuinely complex dispute before a human had already spent time on it. Response times were slow, quality was inconsistent. At the same time, the risk team was reviewing fraud and gambling patterns in transaction data by hand, a process that was slow by design, because humans can only review what they have time to look at. The patterns that emerged across thousands of transactions simultaneously were either caught late or missed entirely.

We built two agents to address both problems at once. The first was an AI Ticket Agent, every ticket that arrives in Zendesk is automatically enriched and classified on arrival, before any human sees it. Simple tickets are resolved autonomously: the agent checks credentials, looks up transaction status, executes the required action, and closes the ticket with the outcome documented. Complex tickets are escalated to the human team, but with everything already prepared, key information highlighted, ticket fields pre-filled from the conversation, context attached. The second was a Fraud and Gambling Detection Agent, an ML-powered system that ingests transaction data, identifies non-human-readable patterns at scale, assigns a risk rating to each case, and surfaces high-priority cases on a live dashboard for analyst review. The risk team no longer reviews transactions manually. They review what the ML model flags as genuinely worth their attention.

What changed was the relationship between volume and headcount. Before Palete, every ticket required a human. After Palete, 70% of tickets are resolved without one, in under a minute, down from hours. The 30% that reach the human team arrive with full context already prepared, which means the team spends their time on judgment calls, not data entry. On the fraud side, the analysts went from manually reviewing a fraction of transactions to having automated coverage across the entire dataset in real time. Over the deployment: 100% of tickets enriched on arrival, 70% auto-resolved with no human needed, 2,000+ tickets processed per month at under 1 minute each, and analyst time freed from manual transaction review entirely.