--- title: "AI agent handoff across channels: what must travel" description: "A handoff works when the customer never repeats themselves. The four things that must travel across SMS, WhatsApp, web chat and voice, and where they get lost." url: https://studiochat.io/blog/traspaso-a-una-persona-entre-canales locale: en translations: { es: https://studiochat.io/es/blog/traspaso-a-una-persona-entre-canales.md, pt: https://studiochat.io/pt/blog/traspaso-a-una-persona-entre-canales.md } site: https://studiochat.io --- [ Back to blog](/blog) A handoff works when the customer never repeats themselves. That requires four things to travel with the conversation: who the customer is, what they asked, what the agent already did and what is left to decide. Across channels the identity is the hardest part, because a phone number, a WhatsApp account and a web session are three different keys. Most handoff problems get diagnosed as a staffing problem. They are usually a state problem: the person who takes over does not know where the conversation stands. ## [Why do handoffs fail?](#why-do-handoffs-fail) Because the agent and the person work in different tools. The agent has a conversation, a few facts it looked up and a list of things it tried. The person opens a ticket, a chat window or a phone, and sees a name and a first line. The customer feels it as a question they have already answered. The cost is not only irritation. Every repeated question is time on the line, and on voice it is time with the customer waiting. ## [What are the four things that have to travel?](#what-are-the-four-things-that-have-to-travel) | What travels | Example | What breaks without it | | -------------------------- | ------------------------------------------------------------------------------ | ------------------------------------------------------- | | Who the customer is | The record that links their phone, their WhatsApp account and their order | The person cannot find the history and starts from zero | | What they asked | "Wrong item delivered, wants a replacement before Friday" | The person asks the opening question again | | What the agent already did | Checked the order, confirmed the address, offered a refund and it was declined | The person offers the refund again | | What is left to decide | A policy exception that only a person can approve | The person does not know why they were called | The last row is the one teams leave out. The agent has not failed when it hands off: it has reached the edge of what it is allowed to decide. Saying where that edge was is the most useful sentence in the handover. ## [What changes when the handoff crosses a channel?](#what-changes-when-the-handoff-crosses-a-channel) The key changes. A customer who started on web chat and is now being called is identified by a session on one side and a phone number on the other. A customer who texted and then wrote on WhatsApp is, to the systems, two people. That is why an agent that orchestrates channels needs a record of its own for the customer, not one per channel. We cover what orchestration has to mean [in another post](/blog/que-significa-orquestar-agentes-de-ia). The handoff is the moment when the shortcut shows, because a person has to read what the agent knew. ## [Does the WhatsApp window affect the handoff?](#does-the-whatsapp-window-affect-the-handoff) Yes, and it catches teams that hand off late. Meta's rule is that free-form replies are allowed inside a 24-hour customer service window that restarts when the customer writes, and once it closes only templates are allowed. A person takes over under the same rule as the agent did. If the handoff happens after the window has closed, the first message has to be a template, and we cover how the [window and templates](/blog/ventana-de-24-horas-whatsapp-plantillas) work in detail elsewhere. ## [What has to travel back from the person to the agent?](#what-has-to-travel-back-from-the-person-to-the-agent) More than teams expect, and the rule that makes it matter comes from consent. When a customer tells a person to stop contacting them, that has to reach the agent. In the US the FCC's revocation rule is explicit, and [a law firm's summary](https://www.bclplaw.com/en-US/events-insights-news/the-tcpas-new-opt-out-rules-take-effect-on-april-11-2025-what-does-this-mean-for-businesses.html) quotes it this way: > revocation extends to both robocalls and robotexts regardless of the medium used to communicate the revocation of consent. The practical reading is that an opt-out said to a person on a call has to stop the agent's texts as well. The [rules by country](/blog/reglas-antes-de-que-un-agente-escriba-o-llame) are in their own post, and in any country the lesson is the same: the record has to flow in both directions. ## [What triggers a handoff?](#what-triggers-a-handoff) Our published description lists four: the person in the conversation asks for one, the sentiment turns negative, the request is a skill the agent does not cover, or the account is one you have flagged as VIP or critical. Each trigger fires at a different moment in the conversation, and the quality of the handover depends on what the agent has written down by then, not on which trigger fired. The handover preserves the conversation history, the data the agent fetched and the actions it attempted, so the person picks up where the agent stopped. That is a description of what we build, and we ask any vendor the same question: show me the screen the person sees when the handoff arrives. ## [How do you know it is working?](#how-do-you-know-it-is-working) Look at the recontact. A handoff that works produces a person who resolves the case without asking the opening question again, and a customer who does not come back about the same problem within the next few days. Count the handoffs where the first human message repeats a question the agent had already asked, and fix those first. Frequently asked questions ### What should an AI agent pass to a human during a handoff? Four things: who the customer is, what they asked, what the agent has already done, and what is left to decide. If any of the four is missing, the person starts by asking the customer to repeat it, which is the failure a handoff exists to prevent. ### When should an AI agent hand a conversation to a person? When the customer asks, when the sentiment turns negative, when the request falls outside what the agent is built to do, and for accounts you have decided always get a person. The trigger matters less than the state of the conversation at the moment it fires. ### Can a human reply on WhatsApp after the 24-hour window has closed? Only with an approved template. Meta's rule applies to a person as much as to an agent: once the 24-hour customer service window closes, free-form messages are not allowed until the customer writes again. A handoff that happens late has to start from a template. Share [ X](https://x.com/intent/post?text=When%20the%20AI%20agent%20hands%20off%3A%20what%20has%20to%20travel%20from%20chat%20to%20voice%20to%20a%20person&url=https%3A%2F%2Fstudiochat.io%2Fblog%2Ftraspaso-a-una-persona-entre-canales) [ LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fstudiochat.io%2Fblog%2Ftraspaso-a-una-persona-entre-canales) [ WhatsApp](https://wa.me/?text=When%20the%20AI%20agent%20hands%20off%3A%20what%20has%20to%20travel%20from%20chat%20to%20voice%20to%20a%20person%20https%3A%2F%2Fstudiochat.io%2Fblog%2Ftraspaso-a-una-persona-entre-canales)Copy link Cristian Pereyra Cristian Pereyra · Co-founder · Engineering, StudioChat Want agents like these working for your team? [Talk to us now](https://cal.com/studiochat)