Skill · July 27, 2026 · 7 min read

The AI escalation ladder: knowing when an agent should ask for help

The most useful agent is not the one that always acts. It is the one that knows which actions are safe, which need approval, and when a person should take over.

A dark visual of an approval surface for AI work

An agent can move quickly while a person still owns the consequential decisions.

Why this matters now

OpenAI’s July 22 announcement for Presence describes agents that can answer questions, resolve issues, use company systems, take approved actions, and escalate to people. That combination is the important design pattern: capability is paired with a route back to human judgment.

The practical question for a smaller team is how to make that boundary explicit before an assistant touches a real inbox, CRM record, calendar, codebase, or customer conversation.

The four levels

1. AnswerLow-risk information with a clear source and no external side effect.
2. DraftUseful work product that a person reviews before sending or saving.
3. Act with approvalA prepared change runs after the owner confirms scope, target, and consequence.
4. Hand offAmbiguous, sensitive, high-impact, or emotionally complex work goes to a person.

How to choose a level

Ask: What can go wrong? Who is affected? Can the action be reversed? A factual answer with a source usually stays at level one. A customer reply becomes a draft when tone or policy matters. A refund, permission change, publication, or deletion belongs at level three or four.

Do not confuse confidence with permission. An agent can be likely to produce the right answer and still lack authority to send it. Confidence is a model signal; permission is a workflow decision.

Copy this workflow brief

Task: [what outcome should exist?]
Default level: [answer / draft / act with approval / hand off]
Allowed action: [exactly what may be read, changed, sent, or saved]
Approval owner: [name or role]
Escalate when: money, access, reputation, legal position, sensitive data, conflicting sources, or scope changes are involved
Evidence: [sources, proposed change, target, and rollback path]

A small case study

Imagine an AI assistant helping a support team with a billing dispute. It can answer a policy question from the approved handbook. It can draft a reply when the customer’s history is relevant. It can prepare a credit adjustment for a manager’s approval. It should hand off when the record suggests fraud, a vulnerable customer, or a policy exception.

Nothing about the ladder makes the assistant less useful. It makes the useful part repeatable: the agent keeps moving until it reaches the boundary, then the boundary is clear enough for a person to make a fast decision.

The takeaway

The best AI workflows do not hide the handoff. They make it a visible feature. Define the ladder once, attach it to the workflow, and review exceptions after real use.

Source

OpenAI, July 22, 2026 — Introducing OpenAI Presence