Progressive Trajectory-Aware Calibration for Human–AI Handoff in Sales Conversations

AI assistants in sales conversations predict conversion probabilities at each turn, but miscalibrated confidence can undermine human–AI handoff decisions. We propose Progressive Trajectory-Aware Calibration (PTAC), a lightweight post-hoc method that adjusts temperature using conversation progress and observed probability trajectories. PTAC requires fewer than 5,000 parameters. On e-commerce data with independently verified outcomes, PTAC reduced turn-stratified calibration error by over 80% at 14% purchase prevalence and by 9% relative to static scaling at the natural 6% prevalence, matching adaptive scaling. It reduced error by 63–78% on call-center data with model-generated labels and by 13% on synthetic dialogues. PTAC improved handoff error rates at fixed human workloads and reduced misplaced reliance at operational thresholds; trust outcomes were simulated and matched-trust performance was comparable. Causal-progress ablation supported deployment-time improvements on call-center and synthetic data; e-commerce gains depended on non-causal turn position.