Policies · AI Escalation Rules
AI Voice Assistant — Escalation Rules
Published for transparency — the trigger words, sentiment signals, context flags, and confidence thresholds that force the AI to hand the call to a human. Hard-coded — no model judgement is allowed to override these. We tune for high escalation rather than low.
The escalation principle
The AI's job is to handle the easy 70% and escalate the hard 30%. We tune for HIGH escalation — better to interrupt a human with an easy case than leave a customer struggling with a stubborn AI. Rules are HARD rules enforced by vendor's intent classifier, hard regex match on speech-to-text, and confidence threshold — NOT model discretion.
Tier 1 — instant transfer (no model judgement)
Request for human: 'human', 'person', 'real human', 'real person', 'real agent', 'speak to a human / person / someone / agent / manager', 'talk to a human / person / someone / agent / manager', 'get me / put me through to / connect me to' + (human / person / someone / agent / manager / operator), 'customer service rep', 'are you a robot? / a bot? / a machine?' + dissatisfaction, 'stop / I'm done / cancel / forget it' + (with the robot / this assistant / the AI), 'I want a manager / supervisor / director / boss'. Subject-matter triggers: 'complaint' → Operations / Director. 'damaged / broken / lost / wrong item' → Claims. 'legal / solicitor / lawyer / sue' → Director. 'GDPR / delete my data / ICO' → DPO. 'sanction / Russia / Iran' + suspicious → MLRO. 'bribery / corruption / fraud' → Director (ethics). 'modern slavery / trafficking' → Director (slavery). 'whistleblower' → Director. Distress / safety: 'hurt / kill myself', 'not safe', domestic abuse indicators, medical emergency → Supervisor + signpost.
Tier 2 — behavioural triggers (auto-escalate)
ASR (transcription) confidence < 0.7 sustained 10s → Handoff. Intent confidence < 0.7 sustained 10s → Handoff. Caller repeats themselves 2 times within 60s → Handoff. Caller's sentiment 'frustrated' 2 times in 90s → Handoff. AI fails to provide a usable answer 2 turns in a row → Handoff. Caller speaks unsupported language → Handoff. Caller's speech rate suggests distress → Handoff. Call duration > 5 minutes (sales) / > 10 minutes (service) → AI auto check-in: 'This is taking longer than usual — would you prefer to talk to someone now?' Caller and AI both speaking over each other 3 times in 30s → Handoff. Caller silent > 12 seconds after a question → one reprompt; if still silent, end politely. Prompt injection: 1 attempt → polite refusal; 2 attempts → Handoff.
Tier 3 — contextual triggers (auto-escalate)
First call from a new caller / number not in CRM: handle greeting + initial query; for booking or commitment → handoff. Caller flagged 'VIP' in CRM: greeting + immediate handoff. Caller flagged 'no AI' preference: greeting only — AI says 'Let me put you straight through to the team' → handoff. High-value or complex shipments: informational queries handled; for booking commitment → handoff. Cargo includes dangerous goods: DGSA review required → handoff for DG questions. Cargo origin / destination is sanctions-jurisdiction: greet + handoff (no information from AI on sanctions questions). Caller is a journalist or PR: handoff to Director (press queue). Caller is from another freight forwarder / industry body: handoff for case-specific. Caller mentions a competitor's brand asking us to comment: polite decline, handoff.
Routing destinations
General enquiries, quotes, bookings (high-value) → Operations / Sales. Tracking, shipment status follow-up → Operations. Claims (damage, loss, delay) → Claims team. Complaints → Operations Stage 1 → Director Stage 2. Billing, refunds, payment disputes → Accounts. Subject access requests, data protection → DPO. Sanctions, KYC, AML → MLRO (Director). Bribery, ethics, fraud → Director (ethics@). Modern slavery, trafficking → Director (slavery@). Whistleblowing → Director. Legal threats, pre-action → Director. Press, PR enquiries → Director (press@). Vulnerable callers → Operations supervisor (extra time). Distress / safety → Operations supervisor + signposting. Language assistance → Operations + translator service.
Warm-transfer summary
Every handoff includes a summary delivered to receiving human (text + brief audio clip option): from (AI assistant + call ID), to (destination queue), reason for escalation (Tier + specific signal), caller (name, phone, customer match, existing flags), call so far (caller intent, what AI told them, what caller said, sentiment, confidence at handoff), suggested next step, link to full transcript. The receiving human reads this BEFORE taking the call so the caller does not have to repeat themselves.
No re-engagement and monitoring
Once handed off, the AI does NOT rejoin the call. If the human needs to look something up, they put the caller on hold normally — the AI does not re-enter. KPIs: Tier 1 trigger missed (post-hoc audit) = 0%. Tier 2 trigger fired correctly ≥ 95%. Tier 3 ≥ 95%. Time to human pickup (during hours) < 30 seconds 95%. Callback within 4 BH for callback-requested 95%. Repeat-handoff rate (handed back to AI) 0%. CSAT at handoff ≥ 4 / 5. 'Robot was unhelpful' complaints per 1,000 calls ≤ 10. Reviewed monthly by Director + DPO + AI vendor. Every change to escalation rules requires Director sign-off, documented reason, sample test against last 100 calls before rollout, version bump, change-controlled vendor deployment. Vendor cannot push rule changes silently — change-control enforced by DPA.
Last reviewed 2026-06-06.