Escalation

​

What escalation does

Escalation is how your AI agent decides “I should stop talking and get a human involved.” When it triggers, Halo creates a ticket in your inbox with the full conversation attached, the user/company context, and any tools the AI tried. Escalation is configured per agent at AI Agents > [Your Agent] > Actions > Ticket Escalation.

​

Modes

Every agent has one of three escalation modes: Most teams start with Conditional to avoid creating tickets for questions the AI could have answered.

​

What signals trigger escalation

The agent looks for three signals on every response:

1. The escalate tool — the AI explicitly calls it because it’s stuck. Always trusted.
2. Escalation phrases in the response — language like “I’ll get a teammate involved” or “Let me create a ticket for this.”
3. Confidence below threshold — when the AI’s confidence score (0–100) drops below your threshold, escalation kicks in.

In Conditional mode, signals 2 and 3 must also pass the gates (knowledge search required, max attempts not yet reached).

​

Configuration options

​

Filters work like segments

Both escalation filters and priority filters use Halo’s segment language — the same one you use to build broadcast audiences and inbox views. You can filter by:

• User traits: name, email, role, plan, custom fields
• Company traits: plan, employee count, MRR, industry
• Stripe data: subscription status, MRR, lifetime value
• Activity: pages visited, last active

For example: “Escalate only if company.plan = enterprise OR user.custom_fields.is_vip = true. When user.custom_fields.is_vip = true, mark as priority.”

​

Manual handoff

Even when automatic escalation is off, users can still ask for a human via the widget — a “Talk to a human” button or by typing escalation phrases. This bypasses confidence checks and gates. You can block manual handoff for specific user segments using escalation filters (when the user doesn’t match the filter, the manual handoff button is hidden). Use this when you want to gate human support behind a paid plan.

​

What happens when a ticket is created

1. Ticket appears in Inbox with status open (or priority if priority filters matched).
2. Conversation history is attached — the full back-and-forth, the AI’s confidence score, escalation reason, knowledge sources searched, tools tried.
3. Notifications fire — Slack (if configured), email (if assigned), real-time inbox updates.
4. Default assignee picks it up — auto-assigned if you’ve set a default; otherwise it sits unassigned.
5. Consolidation — if the user already has an open ticket, new escalations are merged into it instead of creating a duplicate. When the AI is answering an email conversation, escalations and AI-created tickets consolidate into that email ticket itself — the email thread stays the customer’s single conversation, and no separate chat ticket is created.

​

Reading the escalation context in the inbox

Every escalated ticket has an AI Conversation Context card that shows:

• Confidence score (percentage)
• Escalation reason — why the AI handed off
• Conversation history — the full transcript
• Knowledge sources tried — which docs and articles the AI searched
• AI summary — a short summary of what the user wanted

Use this to triage faster — you can see at a glance whether the AI was just stuck on a missing doc (fix the doc) or whether the user has a real edge case.

​

Tuning over time

A good escalation strategy evolves:

• Start in Conditional mode with a moderate threshold (e.g. 50), max 3 attempts, knowledge search required.
• Watch the inbox — for tickets that the AI could have answered, find what knowledge was missing and add it.
• For tickets the AI shouldn’t have escalated — tighten gates or raise the confidence threshold.
• For tickets the AI delayed — lower the threshold or reduce max attempts.

Iterate every few weeks until escalation feels right for your customers.

​

Where to go next