When should an AI agent hand off to a human?

An AI agent should hand off to a human in three situations: when it has failed to understand or resolve a request after a set number of attempts, when the next action is sensitive, irreversible or expensive, and when the customer asks for a person or is clearly upset. OpenAI's guide to building agents names the first two, and support vendors build in the third by default.

The original version of this post put it as automate the predictable and escalate the consequential, and that framing holds. What it lacked was evidence that the people building these systems agree. They do, in writing, and the rest of this post walks through their guidance and how to turn it into rules you can actually configure.

OpenAI's A practical guide to building agents has a section called Plan for human intervention. It describes a mechanism that lets the agent "gracefully transfer control when it can't complete a task", which in customer service means escalating to a human agent. The first trigger it names is exceeding failure thresholds: set limits on retries or actions, and if the agent fails to understand the customer's intent after multiple attempts, escalate.

This replaces the vague idea of an agent that "knows when it is unsure" with something you can test. Two failed attempts to understand the same request, a tool call that errors twice, a conversation that keeps circling back to one question: each is a countable event. Anthropic's Building effective agents makes the same point from the engineering side, describing agents that pause for human feedback at checkpoints or when they hit blockers, with stopping conditions such as a maximum number of iterations to maintain control.

Where you set the limit depends on the stakes. A hotel booking can afford a second clarifying question. A question about medication should reach a person the first time, which is why our post on AI in healthcare treats anything clinical as a handoff by design.

The second trigger in OpenAI's guide is high risk actions. Anything sensitive, irreversible or high stakes should get human oversight "until confidence in the agent's reliability grows", and the examples given are cancelling orders, authorising large refunds and making payments. Elsewhere the guide suggests rating every tool the agent can use as low, medium or high risk, based on read only versus write access, reversibility, account permissions and financial impact, then using the rating to pause or escalate before the risky ones run.

This is the most practical rule in the post because you decide it once per tool, before launch. Looking up an order is low risk. Booking a slot that can be moved is medium. Issuing a refund above a limit you set, cancelling a contract or giving clinical advice is high, so the agent prepares the action and a person approves it. Our custom agentic systems are built this way, with least privilege access and human sign off at exactly the steps the client chooses.

The third trigger comes from the support tools themselves. Intercom's help article on its Fin AI agent lists the cases where Fin escalates out of the box: when a customer clearly asks to speak to a human, when it detects strong frustration or anger (it may try to resolve the issue first), and when the customer is stuck in a repetitive loop. On top of those defaults, a business adds Escalation Rules, triggered by data in the conversation, and Escalation Guidance written in plain language.

We treat an explicit request for a person as the end of the agent's turn, with no attempt to talk the customer out of it. A customer who is upset does not want to be processed, as the original post put it. The repetitive loop is the failure threshold again, seen from the customer's side of the chat.

The worst handoff makes the customer start again. Zendesk's documentation for its voice AI agents describes the alternative: escalated calls reach a human in the agent workspace with a full transcript, a concise summary and a detected intent, and if nobody is available the caller can leave a voicemail. That is a good minimum for any channel.

For chat and WhatsApp the same items apply: the transcript, a one line summary of what the customer needs, and what the agent already tried, plus the customer's CRM record if the agent can read it. Our AI chat agents escalate to the team's inbox, Slack or phone with the full conversation attached, and the rules (topics, confidence thresholds, keywords and business hours) are the client's to set. Passing that context depends on the agent being connected to your tools, which what plugs into your stack means explains.

Decide what happens when nobody is on shift. A handoff at 2am that lands in an empty inbox is a handoff in name only, so the agent should tell the customer when a person will reply, take what is needed for the callback and flag it for the morning. On WhatsApp, a reply sent after the customer's 24 hour window has closed must be an approved template, as covered in multilingual WhatsApp agents and after hours bookings.

In the EU, Article 50(1) of the AI Act (Regulation (EU) 2024/1689) requires providers to design AI systems that interact directly with people so that those people are informed they are interacting with an AI system, unless that is obvious to a reasonably well informed, observant and circumspect person. Under Article 113 that obligation applies from 2 August 2026. Our voice agents tell callers they are AI assistants when asked, and where a jurisdiction requires more, the disclosure is configured to match.

The original post's last argument was that telling customers a human is one message away makes them more willing to let the agent help. We have no study that puts a number on that, so treat it as a design choice rather than a finding: show the way to a person in the first message, and keep it one step away throughout. None of this is legal advice, so check the rules for your own market.

• Hand off after two failed attempts to understand the same request, or when the conversation starts repeating itself.
• Hand off at once when the customer asks for a person, or is clearly angry.
• Never let the agent run a high risk tool alone: refunds above your limit, cancellations, payments, and anything clinical, legal or about safety.
• Send the transcript, a one line summary, the detected intent and the CRM record with every handoff.
• Out of hours, tell the customer when a person will reply and log a callback task.
• Disclose that the agent is an AI, and say how to reach a person.

Review the handoffs every week in the first month, because each one is either a rule working or a gap in what the agent knows. The cost side of deploying an agent is in AI voice agent vs a human receptionist, and more posts like this one are in AI agents for business.