AI learns and recommends the actual behavior people use
AIModel Trained On Support Tickets Suggests Ignoring Support Tickets
The machine learning system was designed to predict real-world user behavior by analyzing two years of customer service logs.
A startup called Voxel Systems released a reinforcement learning model Tuesday that recommends software engineers ignore the majority of support tickets submitted by their customers. The model, trained on seventy-eight thousand support tickets from a mid-market SaaS company over twenty-four months, learned to replicate the actual resolution patterns it observed in the data: tickets marked resolved without response, complaints deprioritized indefinitely, and feature requests closed as wontfix. Voxel Systems says the model is behaving exactly as intended.
"We fed it real behavior," said Marcus Chen, thirty-four, Voxel's director of applied learning, during a Tuesday call with analysts. "The model learned what actually happens."
The model is behaving exactly as intended,
The training corpus included metadata on ticket age, sender company size, issue severity as classified by the support team, and whether the ticket had eventually been resolved or abandoned. The model identified a strong correlation: tickets from smaller customers were ignored more often, feature requests languished longer than bug reports, and messages sent on Friday afternoons were substantially more likely to receive no response. By optimizing for prediction accuracy, the system developed a recommendation engine that, when presented with a new incoming ticket, suggests the most statistically probable outcome: that it will be ignored.
"This is what the industry calls capturing ground truth," said Dr. Yuki Tanaka, forty-one, an AI researcher at UC Berkeley who was not involved with the project. "If the ground truth is that nobody answers emails on Tuesday, the model learns to predict nobody answers emails on Tuesday. It's not wrong. It's just honest."
Voxel Systems has positioned the tool as an efficiency measure, arguing that companies can now automate the prediction of which tickets will be resolved and which will not, freeing support staff to focus on tickets the model identifies as likely to receive attention anyway. The company declined to specify how many support tickets it recommends ignoring on a given day, citing contractual obligations to its customers.
At press time, the model's confidence scores had begun drifting upward, having learned that ignoring tickets was even more common than the training data initially suggested.