Settingsintermediate

Handover to Humans

Handover is the boundary between the Agent Stack and your team. Configure when handover happens, what information must be collected first, and which team picks up — all from the Handover tab.

6 min read

Handover to Humans

Handover is the moment your Agent Stack stops handling a conversation and a human takes over. Sometimes this is graceful — the AI gathers what it needs, summarizes, and a human picks up with full context. Sometimes it’s urgent — the customer is frustrated and the AI gets out of the way. Either way, the Handover tab is where you configure both ends of that handoff.

The three sub-tabs

The Handover tab has three sub-tabs that walk you through the lifecycle of a handover:

  • Triggers — Whether handover is allowed at all
  • Prerequisites — The information the AI must collect from the customer before handing off
  • Delivery — Who picks up, and what happens outside opening hours

Disabling handover at the Triggers level (setting handover mode to “never”) greys out the other two sub-tabs — they only matter once handover is allowed.

Handover modes

A stack can be configured with one of three handover modes (full reference: handover modes):

  • never — the stack never hands off. Conversations either resolve in-AI or the customer gives up. Rare; only sensible for stacks that handle a single closed task with no human-help fallback.
  • by_description — the AI decides when handover is needed based on its trigger rules (frustration, escalation request, stall) and routes to a team derived from the conversation context. A new stack starts in never; turning on Allow handover to humans on the Triggers tab sets this mode.
  • explicit_team — handover always goes to one specific team you choose. Use this when only one team handles all human-needed cases for this stack.

Prerequisites — gather before you handoff

Prerequisites are the questions the AI must have answers to before handing off. Each prerequisite has:

  • A field — either a conversation field or a contact field. The answer is saved into that field.
  • An ask prompt — operator guidance and diagnostic text that tells the AI what information to collect. This is not customer-facing copy: Atender generates the customer-visible ask in the conversation language using the prerequisite field label, so make the field label clear and easy to translate.
  • A condition mode — always (always required) or when (required only when a stated condition is met).
  • A sort order — prerequisites are checked in order.

A typical setup: a billing stack might require the customer’s account email and the invoice number before handoff. The AI checks each prerequisite, asks for anything missing, and only triggers handover once everything applicable is collected. If the customer doesn’t provide a value, the handover does not commit on that turn — what happens next depends on the channel:

  • Generic prerequisite fields on other channels — the gate keeps asking, up to 10 asks per handover attempt, counted across all details. After that, the attempt is given up and the customer gets a full AI turn again.
  • The chat widget’s own email address — this is optional unless an email follow-up needs it, and asked at most twice, counting the one-time-code ask for an address another contact owns. If the customer’s reply to the first ask is a clear refusal, in any language, the next step is the same as after the second ask: the conversation hands over on that reply without the address once nothing else is missing, and the address is not added to the contact. A reply that is not a refusal (for example “hello?” or “why do you need it?”) still gets the second question. After the second ask, the next reply hands over without the address once nothing else is missing, and the address is not added to the contact. A tenant-required field is still asked next after a refusal. The one-time code and email follow-up counts are unchanged. When the handover goes ahead after a refusal, the internal note says the customer did not want to give the address and quotes their reply.
  • A required code — needed for an email follow-up, an always-closed chat, or a workspace field that needs a confirmed email. The code is asked twice. After that, each reply gets a fixed line saying the customer cannot be passed to a colleague without a confirmed address. The code entry stays open, so a later correct code hands over as normal. Nobody is handed over without the confirmed address.
  • Email — the AI asks at most twice per detail. It then hands over without the detail and posts an internal note quoting what the customer wrote after the first question, so your team can see the customer’s own words.

Outside opening hours

The Delivery tab includes opening-hours behavior. When Check opening hours before handing over is on (off by default) and the team is closed, the AI offers the outside-hours options you enabled:

  • Offer a follow-up in this conversation — Tells the customer when the team opens again, and the team replies in the same conversation (default: on)
  • Offer to close and reach out later — Lets the customer close the conversation and try again during opening hours (default: off)

When the team is closed, the reply to the customer names the reopening time in the team’s time zone.

Ask the customer to confirm before handing over (default: off) sits beside these but is not an outside-hours option — it is an independent toggle that is not gated on the opening-hours check — though when the team is closed and an outside-hours option is enabled, that deferral runs instead of the confirmation ask on that turn. When the customer confirms and the team is still closed with the follow-up option enabled, that confirmation is treated as a follow-up; if the team has reopened, the confirmation proceeds as a live handover or queueing. The confirmation question is asked except on email when silent handover is on (see Hand over without a reply on email).

Hand over without a reply on email

This switch is not on the Handover tab. It is Hand over without a reply to the customer (default: off), on the stack’s Personality tab under Channel preferences → Email, in the Behaviour box. It can be turned on only while handover is allowed. It changes what the customer sees when a handover commits on an email conversation. While it is on:

  • A committed handover sends the customer no message, except that a customer who accepts an outside-hours offer still gets a confirmation. The team still receives the conversation and the reason for the handover.
  • The confirmation question is skipped on email, even if Ask the customer to confirm before handing over is on.
  • Other channels are unchanged. Web chat and any other channel behave as described elsewhere in this article.

Silent handover only covers the handover message itself. These messages are still sent to the customer on email:

  • Questions asking for missing prerequisite details. On email, when the AI’s own answer is kept, the question follows that answer and the email ends with a closing; otherwise the question is a short email of its own (greeting, why the detail is needed, the question, and a closing); the greeting and closing follow the stack’s email preferences. The second ask says it is the second request and that the case goes to a colleague without the detail.
  • The outside-hours offer
  • The “handover failed” message

Email follow-up on web chat

What happens when the customer accepts the follow-up on a web chat conversation depends on whether the handover’s team is closed right now, or the chat is always closed, the team is open again, or no team is set.

When the team is closed. The handover commits at once, and the chat stays on web chat. The team receives it with the customer’s email address, and the link Atender emails only confirms that address:

  • The link is valid for 24 hours and is confirmed by a POST, never a plain page load, so a mail scanner cannot confirm it on the customer’s behalf.
  • A reply written after the customer left also goes to the confirmed address, or by SMS to a phone number the customer has already confirmed, when SMS is on.
  • An unopened link does not close the chat. The conversation stays on web chat whether or not the link is ever opened.
  • No link is sent if the customer already has a confirmed email address, or a confirmed phone number while SMS is on, or if the workspace cannot send email.

For an always-closed chat, an open team, or a handover with no team. The existing move-to-email behavior applies. On a web chat conversation, choosing email follow-up waits for the customer to confirm the address:

  1. The handover waits for the link. Atender parks the handover: the AI stays on the chat and the conversation does not enter the team’s queue until the customer confirms. When they confirm, the handover commits.
  2. Moving the conversation onto email waits for proof. Atender parks the channel switch and emails a confirmation link to the address (valid for 24 hours). The conversation only becomes an email conversation once the customer opens that link and confirms — the link is confirmed by a POST, never a plain page load, so a mail scanner cannot confirm it on the customer’s behalf. If the link expires unopened, the conversation does not move to email and the parked handover is dropped: Atender closes the web chat with a localized final line, leaves an internal note for the team, and shows returning visitors the old chat locked with a Start a new chat action.

In this case there is no shortcut: the link is sent even for an address already on the contact. If returning-customer detection finds an earlier conversation, a choice email goes instead of the link and the handover commits on web chat. If the workspace cannot send on email at all, or the widget’s tenant has channel switching unavailable, the switch is skipped and the handover stays a metadata-only handover on web chat.

Handover instructions

The Delivery tab also has a free-form Handover instructions field — extra rules layered on top of the AI’s built-in handover policy: when to hand over, what to gather first, and how to route. Examples: “Mention if the customer has been a customer for over 5 years,” or “Always tell the customer the human’s first name.”

What the human picks up

When handover delivery succeeds, the conversation moves out of the Alia tab and into the chosen team’s queue. The handoff pack the orchestrator’s handover_to_human call produced is attached to the conversation timeline: a summary of what happened, the open question the human needs to settle, the structured fields already gathered from the customer, and the actions the AI already took. The human picks up with all of that available, not from scratch.

For normal configured handovers on non-email channels, if the reachability check fails (Atender cannot assign a team, or handover resolution repeatedly fails), it does not promise the customer a human follow-up. Instead, the AI stays engaged in the conversation and operators are alerted so the delivery problem can be investigated. On email, the AI does not send its “can’t reach a teammate” reply: it stands down and the conversation is handed over to a person instead.

On email, a chosen team that does not resolve (for example “no specific team” or an unknown team key) does not block the handover. The handover still completes and goes to the stack’s default team if it has one, otherwise to the conversation’s current team. If neither applies, the conversation lands unassigned and owners and admins get a “Handed over, no team assigned” notification. With silent handover on, nothing is emailed to the customer in this case.

Repeated guardrail failures are not a handover. When a reply fails the output guardrails twice, the platform never forces a handover in any mode: it ships the canned “ran into a problem, try again” line, records the turn’s outcome as a forced fallback, and leaves the AI engaged so it can try again on the next turn. A handover the discarded reply itself asked for is not dispatched either, because the whole turn was thrown away. If the conversation genuinely needs a person, the orchestrator hands over on a later turn through the normal path above.

See also

Tags

Ai FeaturesConcept