AI Still Can’t Beat the On-Call Engineer: Here’s Why

AI corporations hold pitching autonomous site reliability engineer agents—AI that investigates manufacturing incidents instead of people. Datadog ran the precise benchmark on actual outages, and the very best AI fashions cannot but beat the engineers they’re supposed to interchange.

The benchmark is ARFBench (Anomaly Reasoning Framework Benchmark), a joint mission from Datadog and Carnegie Mellon. Constructed from 63 actual manufacturing incidents, extracted from engineers’ personal Slack threads throughout stay emergencies—750 multiple-choice questions overlaying 142 monitoring metrics and 5.38 million knowledge factors, each query verified by hand. No artificial knowledge. No textbook eventualities.

“Trillions of {dollars} are misplaced annually on account of system outages,” the researchers write. The benchmark exams whether or not AI can truly assist change that.

“Regardless of the central position of such question-driven evaluation in incident response, it stays unclear whether or not fashionable basis fashions can reliably reply the sorts of time collection questions engineers ask in apply,” the paper reads.

Questions are available in three tiers. Tier I: Does an anomaly exist on this chart? Tier II: When did it begin, how extreme is it, what sort?

The Tier III—the toughest—requires cross-metric reasoning: Is that this chart inflicting the issue in that different chart? That is the place AI falls aside. GPT-5 scores simply 47.5% F1 on Tier III questions, a metric that penalizes fashions for gaming solutions by choosing the most typical class.

“Regardless of the central position of such question-driven evaluation in incident response, it stays unclear whether or not fashionable basis fashions can reliably reply the sorts of time collection questions engineers ask in apply,” the researchers write.

How each mannequin stacked up

GPT-5 led all present fashions at 62.7% accuracy—on a take a look at the place random guessing will get 24.5%. Gemini 3 Professional scored 58.1%. Claude Opus 4.6: 54.8%. Claude Sonnet 4.5: 47.2%.

Area consultants scored 72.7% accuracy. Non-domain consultants—time collection researchers at Datadog with out in depth observability expertise—nonetheless hit 69.7%.

No AI mannequin beat both human baseline.

The mannequin that truly topped the total leaderboard was Datadog’s personal hybrid: Toto—their inner time collection forecasting mannequin—mixed with Qwen3-VL 32B. Toto-1.0-QA-Experimental scored 63.9% accuracy, edging previous GPT-5 whereas utilizing a fraction of its parameters. On anomaly identification particularly, it outperformed each different mannequin by no less than 8.8 share factors in F1.

A purpose-built area mannequin, skilled on observability knowledge, outperforming a frontier general-purpose system at this particular process is the anticipated end result. That is the purpose.

Probably the most worthwhile discovering is not which mannequin scored highest.

“We observe considerably totally different error profiles between main fashions and human consultants, suggesting that their strengths are complementary,” the researchers write. Fashions hallucinate, miss metadata, and lose area context. People misinterpret exact timestamps and infrequently fail on complicated directions. The errors barely overlap.

Mannequin a theoretical “Mannequin-Professional Oracle”—an ideal decide that at all times picks the proper reply between the AI and the human—and also you get 87.2% accuracy and 82.8% F1. Method above both alone.

That is not a product. It is a documented target—constructed from actual emergencies, not curated datasets—that quantifies precisely how a lot better human-AI collaboration may carry out. The leaderboard is stay on Hugging Face. GPT-5 sits at 62.7%. The ceiling is 87.2%.