Background: Real-time surgical computer vision (CV-AI) is technically feasible, although real-world actionability and failure modes remain poorly characterised. This study quantified live performance and mapped dominant failure mechanisms during intraoperative deployment in laparoscopic cholecystectomy (LC). Methods: A real-time multi-task CV-AI was implemented prospectively (surgeon-blinded) during 100 consecutive LCs. Live video inference evaluated (1) five-phase workflow recognition, (2) inflammatory grade (1-4), (3) Rouviere’s sulcus (RS) identification, and (4) drain detection performance. Outputs were compared with intraoperative observer annotations. Discordant cases underwent structured observer/surgeon feedback integrated with quantitative error signatures to derive a task-by-mechanism failure taxonomy. Results: Phase recognition showed greatest potential actionability. Across 316 eligible transitions, median absolute timing error was 22s (IQR 9-91); only 55% met the predefined ±30 s actionability threshold. Large delays clustered around branched workflows, particularly IOC/CBDE (median error 588s vs 23s). Inflammatory grading showed modest ordinal agreement (κw 0.39), improving when restricted to a post-liver-lift assessment window (κw 0.47). RS outputs were dominated by early false positives (sens 100%, spec 0%). Drain detection performance supported automated documentation (sens 90.9%, spec 83.1%). Mixed-methods synthesis identified unstable/limited visual access, non-linear workflow sequences, instrument-driven proxy triggering, and documentation-reference discordance as predominant failure mechanisms. Conclusion: Real-time CV-AI can generate clinically-relevant intraoperative signals, but utility is constrained by timeliness and predictable context-dependent failure modes rather than headline accuracy alone. Mapping failure mechanisms to quantitative signatures identifies actionable targets for iterative model redesign, including branch- and uncertainty-aware workflow modelling, before progression to higher-stakes decision-support.