PLACO Framework Boosts Human-AI Teams with Cost-Effective Classification

A new research paper from Pranavkumar Mallela, Vinay Kumar, Shashi Shekhar Jha, and Shweta Jain presents PLACO, a multi-stage framework designed to make human-AI teams more cost-effective. The framework addresses classification tasks where a single hard label is the final output. PLACO combines a deterministic human labeler with a probabilistic AI classifier using Bayes rule, crucially assuming that human and model outputs are conditionally independent given the ground truth. It uses the model's instance-level calibrated probabilities and the human's class-level calibrated probabilities to optimally fuse decisions, reducing errors while keeping human effort low.

As generative AI becomes ubiquitous in tasks like writing and algorithm development, PLACO offers a principled way to allocate work between humans and models. By formally modeling the reliability of each party, the framework can dynamically decide when to rely on the AI, when to consult a human, and how to combine both. This approach promises to accelerate workflows without sacrificing quality, making it especially valuable for high-stakes classification in healthcare, finance, or content moderation. PLACO provides a mathematical foundation for building smarter collaboration systems that are both accurate and budget-conscious.