Agent Aid

When to Route an AI Workflow to a Real Person

Decision heuristics: tail risk, novelty, missing tools, legal exposure, and customer trust—plus how to measure when human routing pays off.

Routing everything to humans defeats automation; routing nothing invites incidents. Good systems use a small set of durable rules plus metrics that prove when automation is safe to expand.

Route when tail risk is asymmetric

If a mistake is rare but catastrophic—wrongful account closure, unsafe medical advice, fraudulent transfer—default to human review even if the model is usually right.

Route on novelty

Attacks and edge cases cluster in long tails. The first time you see a pattern, humans label it; later you can automate with tighter detectors.

Route when tools are missing or broken

Agents should not silently improvise. If an integration returns an unknown state, pause and assign a task with the raw error and customer identifiers (redacted as needed).

Route for trust moments

Sometimes the customer simply needs a person—even if the bot could handle it. Offering a human path protects NPS and reduces chargebacks. See customer trust & safety.

Measure the layer

Track human task volume, time-to-resolution, and rework rate. If humans frequently disagree with the agent, fix upstream prompts or tools—not only the reviewers.