← News

Models & toolsSources: arXiv · 18 Nov 2026

Study finds decision models can match frontier LLMs on contact center classification at lower cost

A new arXiv paper compares nine decision models against thirteen generative LLMs on eight contact center tasks, reporting comparable classification performance at lower per-request cost.

In plain words

Contact centers must make bounded choices, like identifying why a caller is calling or grading AI outputs. Decision models return a typed answer with a probability instead of generating text, so they can be cheaper. The paper reports many matched frontier LLMs on classification tasks.

What was announced

The authors compared nine decision models with thirteen generative LLMs in zero-shot settings across eight contact center use cases: six classification tasks and two judging tasks. They report that many decision models matched the performance of frontier LLMs on classification, and that API-based closed decision models did so at the lowest per-request cost among the API models evaluated. Confident predictions from decision models were also more accurate, which the authors say enables selective routing in production.

The study also identifies weaknesses. Across classification tasks, most decision models rarely chose the no-answer option, and the authors attribute most of their gap to the strongest LLMs to this abstention behavior, with Gemini-Decision-1 the notable exception. On judging tasks, quality dropped sharply for smaller decision models, which were also more prone to position bias. The authors frame their results as guidance for choosing models based on abstention behavior, confidence, and deployment cost rather than accuracy alone.

What it does not show

  • The evidence is the paper's abstract only; full experimental details, datasets, and exact figures are not available in the supplied passages.
  • The findings come from the authors' own experiments and have not been independently confirmed.
  • The abstract does not name the specific models, tasks, or per-request costs beyond the exception Gemini-Decision-1.

For security and IT teams

Scope
The paper covers bounded contact center decisions: intent identification, knowledge-base article retrieval, agent compliance assessment, and judging AI model outputs, in zero-shot settings.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.