Gives a way to decide which agent actions need human review by comparing error cost with reviewer time in dollars, instead of using intuition. Useful when designing escalation rules for agentic workflows.
Key takeaways · AI-distilled
Scale's rule for human review: escalate an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → decision when (1 - confidence) x severity x credit value exceeds review cost, so the confidence needed for autonomous action rises with the dollar impact of an error.
A severity factor separates error types: a credit wrongly closed is silently lost in full (severity near 1), while a wrongly posted credit gets disputed and reversed, so it can carry higher values before review.
At 92% accuracy, a $150 review cost and full severity, the agent would escalate credits above $1,875; at 98% confidence the threshold is $7,500, and at severity 0.1 it is $18,750. Scale calls these figures illustrative.
Because LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → self-reported confidence is overconfident, the system splits confidence into signals for reading source data, retrieving records and reasoning, then calibrates the combined score so 0.9 means about 90% correct.
Under a fixed review budget, ranking decisions by expected loss and spending analyst minutes in that order sets the threshold where the budget runs out. In a month of testing, the agent found several times more credits than the old process.
Terms in this piece · Glossary
human-in-the-loop — Requiring a person's approval at specific points in an automated process, chosen so the irreversible steps are the ones a human sees.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
