---

Escalation Intelligence

TL;DR

• AI can flag tickets for escalation: complexity, sentiment, account value, or pattern matching. It will also over-escalate (wasting engineering) and under-escalate (losing customers).
• Use AI for signals. You define the rules. You own the judgment for edge cases.
• Escalation is a balance. Too many: eng is flooded. Too few: customers churn. Tune carefully.

When should a ticket go to engineering? Or to a senior support rep? Or to the customer success team? AI can suggest based on keywords, sentiment, or similarity to past escalations. It can also cry wolf or miss the quiet crisis. Your job: use AI for input, own the escalation policy, and refine based on outcomes.

What AI Can Signal

Complexity indicators:

• Technical terms, error messages, stack traces. AI can flag "likely needs engineering."
• Useful. Verify. Some "complex" tickets are FAQs with extra detail. Some simple-looking tickets need engineering. Human review for borderline.

Sentiment and urgency:

• "Frustrated," "angry," "third time." AI can detect. Good proxy for "escalate sooner."
• Don't auto-escalate on sentiment alone. Some angry customers have simple fixes. Some calm customers have critical bugs. Combine with other signals.

Account and history:

• High-value account. Previous escalations. Churn risk. AI can flag from CRM or ticket history.
• Critical. These should rarely wait. Human routing for these.

Pattern matching:

• "This looks like the bug we escalated last week." AI can surface similar past tickets. You decide if it's the same. You escalate if needed.
• AI suggests. You confirm. Don't auto-escalate on pattern alone — might be a different issue.

What AI Gets Wrong

Over-escalation:

• AI is conservative. "Might need eng" becomes "always escalate." Engineering gets flooded with tickets that support could have solved.
• Tune the threshold. Measure: What % of AI-suggested escalations were actually needed? Reduce noise.

Under-escalation:

• Quiet, polite customers with critical issues. AI might miss. "Seems fine" but it's not.
• Have a safety net: human review for high-value accounts, or a "when in doubt, escalate" rule for certain segments.

Context blindness:

• AI doesn't know the internal roadmap. "We're fixing this next sprint" — don't escalate to eng. Route to CS or leave with support. AI can't make that call.
• You need policy. You need updates. AI suggests; you apply context.

Designing the Escalation Flow

1. Define tiers — What goes to L1, L2, eng, CS? Document. AI can suggest; you assign.
2. Set thresholds — Confidence, sentiment, account. Tune based on outcomes. Too many escalations? Raise the bar. Too many misses? Lower it.
3. Human override — Always. Support rep can escalate even when AI says no. Support rep can de-escalate when AI says yes. Human has final say.
4. Feedback loop — Track: Was this escalation correct? Use that to tune. Monthly review.

Your Role

• Policy owner. You (or your team) define when we escalate. AI assists. You decide. Update as the product and org change.
• Calibrator. You see the tickets. You know when AI is right and when it's wrong. You provide the signal to improve.
• Bridge. You're the one who talks to customers and eng. You translate. AI doesn't do that. You do.

Do This Next

1. Audit last month's escalations — How many were necessary? How many could have been solved at L1? What did we miss? Use that to tune.
2. Define "always escalate" and "never escalate" — Explicit rules. Document. Implement. Reduce ambiguity.
3. Add one safety check — e.g., "All tickets from top 10 accounts get human review before closure." Whatever fits your context. Protect high-value.