When should an AI agent hand off to a human?

An AI agent works inside a defined scope. It can look up approved information, ask questions, classify an intent or start an action. That does not mean it should make decisions about every situation.

Channels have rules too. On WhatsApp, Meta documents webhooks for receiving messages and delivery statuses, and treats business initiated messages through templates differently. Phone and text have their own requirements. The agent's design has to account for these limits from day one, not after a reply goes wrong.

These signals are a starting point, not a universal formula. A law firm, a clinic, an online store and an agency will set different thresholds. What matters is that your team can explain why a conversation stayed automated or ended up with a person.

In sales, a rule might say: if the lead asks for a custom proposal, mentions a budget, wants to negotiate terms or asks for a meeting, the agent collects the minimum details and hands off to the sales owner.

In support, a rule might say: if there is a repeat complaint, a refund request, an error that is not in the approved knowledge base or a customer who is frustrated after one attempt to fix the issue, hand off with priority and a summary.

The rule has to be observable. Instead of "if they seem angry", use signals you can review: keywords, the number of retries, the classified intent, no match in the knowledge base or an explicit request to talk to someone.

The person taking over should not get a lone screenshot or a vague label like "complex inquiry". They should see a short, actionable summary:

• The main reason for contact and what the customer wants.
• Name, phone number and any other details the customer agreed to share.
• A summary of the relevant messages, not the whole chat.
• The answers and sources the agent already used.
• Missing details and the specific question still open.
• Priority, suggested owner and next step.

Separate three things: what the agent can tell people, what it can start and what only a person can approve. For example, it can share your hours and it can log a request, but it may not be allowed to confirm a discount, commit to a date or change a contract.

Keep its sources narrow and current. If there is no approved answer, the safe move is to admit the limit and offer a handoff. Making up an answer to avoid escalating usually shifts the cost to your team and your customer.

Do not just count how many conversations were handed off. Check whether each handoff reached the right person, how much context came with it, how long the first human reply took and whether the customer had to repeat anything.

Compare outcomes too: resolved, qualified lead, meeting requested, complaint closed or conversation abandoned. A higher handoff rate can be an improvement if the agent used to hold on to risky cases. A lower rate can be bad news if it is hiding mistakes.

• Define the scope and the actions the agent cannot approve.
• Write observable escalation signals for sales and support.
• Assign an owner, working hours and a channel for each type of handoff.
• Design the summary the person will see when they take over.
• Test normal, unclear, sensitive and out of scope cases.
• Measure handoff quality, pickup time and final outcome.
• Decide how to pause or fix the flow if errors show up.