Escalate

Pause the AI on this thread and alert a person — the safety net when the AI could not, or should not, finish a path alone.

Quick start

  1. Go to Automate → Workflows and open a workflow. Click + and pick Escalate.
  2. There is nothing to configure — it pauses the AI and alerts a person now.
  3. Add an Assign To (human) step before Escalate, so the alert reaches one person, not every owner and admin.
  4. Set Status to Published in the top toolbar, Save, confirm Publish. Agents: ask an owner or admin.

Watch out

  • An AI app anyone on the team connected under Settings → Profile → Connected apps can still message the customer after this — it ignores this pause.
  • Escalate shows as done even if nobody actually received the alert — do not rely on it as your only way to know.
The Add Steps menu open on the canvas, with Escalate highlighted.
The Escalate step panel, which has no fields — it just flags a human and pauses AI.
The Escalate step placed on the canvas.
Details

Getting here

Automate → Workflows, open a workflow, click + on the canvas, then pick Escalate under Add Steps.

Before you start

  • An owner, admin or agent can add this step and save the workflow. Publishing it is an owner or admin action.
  • Building and running workflows depends on your plan; see pricing.
  • Nothing outside Inchat has to be connected, and there is nothing to configure — Escalate takes no fields.

Set up

  1. Open the workflow and click + where the escalation belongs.
  2. Pick Escalate. There is nothing to configure.
  3. Save the step. To publish: set Status to Published in the toolbar at the top of the editor, click Save, then confirm with Publish. The dropdown on its own changes nothing until you save.

How it works

  • The step writes an internal-only note on the conversation's timeline, recording that a workflow escalated it — the customer never sees this note.
  • It records the escalation alongside any the AI decides on its own, so both show up in the same place for review.
  • It hands the thread to a person the same way an inbound "let me speak to a person" does: the conversation becomes a person's (Human replying) and the AI stays quiet.
  • That pause is read by the customer's next inbound message and by an AI Reply step in this run or one it hands off to.
  • But it is not read by an AI app connected over MCP — that path checks only the separate hold Pause AI / Unassign sets, not this one.
  • So an MCP-connected app can still message the customer on a thread Escalate just silenced.
  • This is a known gap, not something this step is meant to prevent today.
  • Escalate may also add an AI-written summary note and a next-day task for the assignee — but not if the AI already wrote a summary since your team's last reply.
  • An AI Reply step's own decline already writes one of these summaries, so an Escalate step right after it usually adds neither.
  • It alerts a person: a push notification through the same channel any "a human is needed" moment uses.
  • The alert follows each person's own notification settings — their scope, mute setting and at-desk status.
  • It also sends an in-app notification, addressed to the conversation's assignee if it has one and can still reach the channel, or to every owner and admin if it does not.
  • It does not assign the conversation to anyone — a thread can be escalated and still show as unassigned.

When it fails

  • The alert can fail to reach anyone — a push and an in-app notification can each fail independently.
  • Today the step is still recorded as having succeeded even when neither one lands.
  • So there is no separate signal in the run log for an alert that reached nobody.
  • Two of this step's own writes can fail it with no reason text attached — the pause itself, and recording the escalation for review.
  • So the run log can show the step failed without saying which of the two.
  • A third cause names itself, though: the thread moved out from under the step while it was running — for example, a person took it over at the same moment.
  • That stops the run rather than continuing past it.

Limits

  • It sends the customer nothing — both the note it writes and the drafted handoff summary are internal only.
  • After it, this run sends the customer nothing more — not even after a Wait.
  • Its later message, question, AI Reply and Trigger Another Workflow steps are logged as not run, and a Wait ends the run.
  • Tags, fields, status and assignment steps after it still run, unless a Wait comes first — except Assign To (AI), which is logged as not run: the thread stays a person's.
  • It does not assign the conversation to a person; Assign To (human) is what places it in someone's queue.
  • The summary note and follow-up task are both skipped if this stretch of the conversation already has a summary.
  • The task additionally needs the conversation to already have an assignee — it goes to them, never a new pick — and the AI agent's own follow-up-task setting to be on.
  • A person's reply does not give the thread back to the AI: it stays a person's until someone clicks Return to AI.
  • Until somebody replies, the thread is listed in the Inbox's AI handoff view, and the reminders for conversations waiting on a person cover it.
  • On an escalated thread the Inbox offers Return to AI in place of Take over on the web, and Hand to AI in the app.
  • Running an Assign To (AI) step after an escalation does not lift this hold, because that step never clears it.
  • Later in the same run that step is logged as not run, and the AI stays silent.
  • A different automatic run meeting the hold is refused and stops there.
  • It sets a different hold from the one Pause AI / Unassign sets — clearing one does not clear the other.
  • A failed alert is not distinguishable from a successful one in the run log today — the step succeeds either way.
  • The MCP send_message tool does not check this hold at all — see How it works.
  • Treat Escalate as silencing Inchat's own auto-reply and workflow paths, not every way an AI-authored message could reach this thread.

Best practices

  • Assign the conversation to a person before escalating it, not after.
  • Put any message the customer should get, such as "a colleague will reply soon", before Escalate: nothing after it is sent.
  • Escalate's own alert is addressed to the conversation's existing assignee — or to every owner and admin if there still isn't one.
  • So assigning first is what gets the alert to a specific person, rather than the whole team.
  • Reach for Escalate when the point is *a person needs to look at this now*. Use Assign To (human) or Assign To (team) when the point is simply *this is theirs* with no urgency implied.
  • Do not assume Escalate always adds its own follow-up task after an AI Reply decline in the same path.
  • The decline itself already writes a summary, so an Escalate step right after it typically creates neither a second note nor a task.
  • Add a Create a Task step explicitly if the flow needs one there.

Use cases

  • An AI Reply step declines to answer; the path assigns the conversation to a person, then escalates.
  • The AI's decline already alerted whoever holds the thread; what Escalate adds on top is the AI pause and an alert addressed specifically to that assignee.
  • A customer's message matches a Condition built to catch a complaint, and the flow escalates immediately rather than letting the AI attempt a reply.

FAQ and troubleshooting

Does the customer see that the conversation was escalated?
No. The note it writes is internal only, visible to the team, never to the customer.
Does Escalate assign the conversation to someone?
No. It alerts people and pauses the AI, but the conversation can still show as unassigned afterward. Pair it with an assignment step if it should also land in a specific queue.
How do I know if the alert actually reached someone?
The run log cannot tell you today — the step is recorded as succeeded even if both the push and the in-app notification failed to reach anyone.
Does Escalate create a task too?
Sometimes. It can draft a second, AI-written summary note and a next-day follow-up task. Both are skipped if this same stretch of the conversation — since the last human reply — already has a summary; an AI Reply step's own decline writes one, so Escalate right after it usually creates neither. The task specifically also needs the conversation to already have an assignee, and needs the AI agent's own follow-up-task setting to be on; the note does not depend on either of those.
Can an AI app connected over MCP still message the customer after Escalate?
Yes, today it can — the MCP send tool checks a different hold, the one Pause AI / Unassign sets, not the one Escalate sets. This is a known gap on the product side, not something this step currently prevents.

Reading this with an AI assistant? Plain Markdown version.