AI routing of technical support tickets costs five times human sorting

Sending a single technical support ticket to the engineering team incorrectly costs five times what its manual sorting by a human employee initially requires. This computational paradox is the core of a new research paper presented by researcher Sashin Gupta, accepted at the 38th International Conference on AI Tools 2026, which highlights an economic trap ignored by hasty automation projects; misrouting reports consumes the entire automation surplus and doubles operational waste.

The paper introduced an evaluation framework based on the principle of cost-aware abstention, an offline assessment that measures system readiness before actual deployment by freezing a routing policy previously calibrated on calibration data, then testing it on isolated data. The researcher tested this framework on a public dataset comprising 47,837 technical support tickets distributed across eight categories, and compared four conventional classifiers not based on generative models with four large language models operating in a zero-shot and few-shot prompting style.

The study relied on an illustrative cost model that assigns a cost of zero to correct routing, one unit to human verification, and five units to incorrect routing.The practical result showed that simple statistical algorithms crushed generative models in cost control.The traditional logistic regression classifier combined with a term-frequency technique, at a confidence threshold of 0.60, reduced the test cost from 0.763 to 0.472, a relative decrease of 38.1 percent, while cutting the misrouted ticket rate from 15.3 percent to just 3.5 percent, outperforming the always-route and always-abstain baselines in 14 out of 16 test cases.

The technical gap became evident when a stratified test sample of one thousand tickets was examined; the strongest generative model achieved an accuracy of only 0.477, failing to surpass the 48 percent correct-assignment threshold. In contrast, a miniature DistilBERT model trained on the same data reached an accuracy of 0.879, nearly double the performance of all prompts directed at large models, demonstrating that size and general linguistic capability do not compensate for task-specific training.

The paper revealed a structural flaw in the confidence behavior of large language models when dealing with economic decision thresholds. Testing GPT-4 or Mini in few-shot mode showed a single repeated confidence value before and after recalibration, recording an area under the operating characteristic curve for prediction correctness of exactly 0.500, a result equivalent to pure random guessing. Applying an analytical break-even threshold of 0.80 led seven generative settings to abstain from routing any tickets, referring them all to humans, whereas the Haiku model dared to route only 19 tickets out of one thousand.

We saw in this analysis a direct indication for IT departments in Gulf banks, telecom companies in Egypt, and service institutions in the Levant, which are rushing to replace primary sorting offices with generative AI interfaces.Calculate the cost of error before calculating the cost of a programmatic call;A model that misclassifies a server or network outage report and forwards it to application developers instead of operations engineers wastes costly manual re-routing hours that exceed the salary of a human sorter by several multiples. The most engineering-wise and financially sound approach for regional teams is to build lightweight, specialized local classifiers with calibrated thresholds, rather than squandering budgets on giant models that hesitate at decisions or route reports down the wrong path.