Policies · Human Handoff
Human Handoff Policy
You can ask for a real person at any time during an AI call — we transfer you straight away, no questions, no delays. The AI also escalates automatically when it detects you need a person (complaints, claims, frustration, sensitive topics, or when it just cannot help). Whenever we transfer, the human gets a short summary so you do not have to repeat yourself.
The principle
The right to a human is absolute and immediate. The AI does NOT: gate the transfer behind verification, delay to 'try one more thing', insist the caller explain why they want a human, imply the human will be unavailable, slow, or expensive.
Transfer on request — instant
Transferred immediately when caller says: 'human', 'person', 'real human', 'real person', 'real agent', 'speak to / talk to / get me / put me through to' + (human / person / someone / agent / manager / operator), 'customer service rep', 'are you a robot?' + dissatisfaction, 'stop / I'm done / cancel / forget it' + (with the robot / this assistant / the AI), 'I want a manager / supervisor / director / boss' — and the many natural-language variants the AI is trained to recognise.
Transfer on detection — automatic
Complaint language ('complaint', 'unhappy', 'unacceptable') → Operations / Director. Claim language ('damaged', 'missing', 'lost', 'broken', 'wet', 'wrong item') → Claims. Legal language ('legal', 'solicitor', 'lawyer', 'sue', 'court', 'tribunal') → Director. Refund / payment dispute ('refund', 'chargeback', 'dispute the charge') → Accounts. Sanctions / KYC sensitivity → MLRO (Director). Data protection ('GDPR', 'data subject', 'delete my data', 'subject access', 'ICO') → DPO. Safety / emergency → human + signpost 999 / Samaritans. Vulnerable indicators → human supervisor. AI low confidence: <0.7 sustained 10 seconds → human. AI cannot help (2 unrecognised intents in a row) → human. Caller repeats themselves twice → human. Caller expresses frustration twice (detected sentiment) → human. High-value or sensitive shipment → human. Bookings above threshold → human. First-time caller from new entity → human (post-greeting). CRM flagged 'VIP' or 'human always' → human.
Warm transfer — the standard
Every handoff is WARM: the AI transfers with a context summary so the human knows: caller's name (if known), customer (if matched), reason for call, what the AI did / answered, why escalated, sentiment, sensitive flags, suggested next steps. The summary is shown to the human BEFORE they take the call so they can open with the brief. We do not make the caller repeat themselves.
Handoff line
'Of course — putting you through to a member of the team now. They'll see what we've talked about so you won't have to repeat yourself. One moment.' Variants allowed; mandatory elements: polite acknowledgement, statement of transfer, reassurance that context is being passed (no repetition), time expectation.
When queue is full / out of hours
If no human available immediately: AI tells caller — 'The team are with other customers. Two options — wait about 5 minutes, or I can schedule a callback within 4 hours. Which works?' Callback or wait per caller choice. For complaints, claims, DPO requests: callback within 1 business day MAXIMUM. Outside business hours: routine enquiries — AI handles or schedules callback. Complaints, claims, legal threats — AI captures fully, marks high-priority, emails Director-on-call. DPO requests — captures and emails dpo@; DPO responds next business day per UK GDPR 1-month window. Emergencies — AI signposts 999 and Samaritans (116 123).
Special routes
Crisis routing (distress / crisis indicators): empathy, signpost 999, 116 123 Samaritans, 0808 2000 247 National Domestic Abuse Helpline as appropriate; transfer to human supervisor immediately; log for supervisor review same day. Vulnerable adult / safeguarding: transfer to human, may consult Director. Compliance hot-routes: 'I want to report bribery / corruption / fraud' → [email protected] queue. 'Modern slavery / trafficking' → [email protected] queue. 'Suspicious Activity Report' → MLRO with NCA SAR signposting. Once handed off to a human, call stays with the human — AI does not re-join mid-call.
Last reviewed 2026-06-06.