Switchboard detects vulnerable-customer signals in real time across every channel and hands the conversation to a trained human handler with full context. Designed to support — not replace — your vulnerable-customer process.
Hard handoff. Human judgement. Configurable thresholds. Channel-agnostic.
Vulnerability signals are a separate, harder escalation path than confidence or frustration. The AI does not finish the workflow first; it transfers immediately to a human handler with the full conversation context.
Switchboard does not classify a customer as vulnerable on the platform record — that remains a human judgement, made by your trained vulnerable-customer team. The AI surfaces the signals; the human assesses and decides.
Sensitivity thresholds and signal definitions are configurable per agent definition. Insurers with mature vulnerable-customer policies can encode their existing signal library directly. Pilots typically tune over the first two weeks.
Same detection logic on voice, chat, WhatsApp, SMS, and email. Voice handoffs are warm transfers; async-channel handoffs route the thread to the vulnerable-customer queue with the same context summary.
Four Signal Categories
Specific phrases and patterns are configurable — most insurers encode their existing vulnerable-customer signal library into the agent definition.
Imminent harm, unsafe accommodation, threats, accident-with-injury, or health crisis disclosures during the conversation.
Examples
Disclosed medical needs, mental-health markers, recent bereavement, or impacts of disability on the customer's ability to manage the claim.
Examples
Tone, pace, repetition, and language patterns that indicate the customer is in genuine distress regardless of explicit content.
Examples
Indicators that the customer's circumstances change how the claim should be handled — children, elderly relatives, sole income, accessibility needs.
Examples
The Handoff
Vulnerable-customer handoff is faster and harder than the standard escalation path. The AI does not finish the workflow first.
Switchboard detects a vulnerability signal — explicit disclosure, distress marker, life-safety language, or capacity indicator. Signal classification logged in Timeline with the relevant turn.
On voice, the AI transitions to handoff within the same turn — no further questions, no workflow steps. On async channels, the thread is routed to the vulnerable-customer queue immediately.
Receiving handler gets: conversation transcript, detected signal type, partial case data, prior interactions, and any flagged vulnerability history. No cold transfer, no lost context.
Trained vulnerable-customer handler picks up the call or thread. They make the assessment, follow your VC procedure, and update the case record. The AI's role ends; human judgement starts.
Consumer Duty Alignment
Consumer Duty, FG21/1 vulnerable customer guidance, FOS evidence — the same mechanism produces all three.
Consumer Duty requires firms to avoid causing foreseeable harm to customers. Detecting vulnerability signals during AI conversations and routing to a human is foreseeable-harm avoidance by design.
FCA guidance on vulnerable customers expects firms to identify vulnerability, understand the customer's circumstances, and adapt service accordingly. The AI surfaces; the human adapts. The mechanism mirrors the guidance.
When a FOS file is opened on a vulnerable-customer case, the evidence pack shows the signal that was detected, the handoff that followed, and the human handler's actions. Defensible end-to-end.
Aggregated reporting on signals detected, handoff times, and handler outcomes. Vulnerability-detection rate becomes a measurable operational metric rather than an unmeasurable compliance hope.
FAQ
No. The AI detects signals and routes the conversation to a trained human handler. The vulnerability assessment itself — and the decision to flag the customer's record as vulnerable — is a human action by your trained vulnerable-customer team. Switchboard supports the process; it does not replace human judgement.
Four signal categories: life-safety language, medical/health disclosures, distress markers (tone, pace, repetition), and capacity/dependency indicators. Specific phrases and patterns are configurable — most insurers encode their existing vulnerable-customer signal library into the agent definition during onboarding.
Like any detection system, Switchboard's signal detection has a false-negative rate. Mitigation runs on multiple layers: configurable sensitivity (lower thresholds catch more), customer-explicit-request escalation always works (a customer asking for a human is immediate handoff), and the handler picking up after the AI also has signal-detection responsibilities. The AI is one safety net among several, not the only one.
False positives — non-vulnerable customers routed to the VC team — are operationally cheap. The receiving handler completes the conversation either way; the only cost is queue routing. We tune sensitivity during the pilot to balance detection rate against handler queue load. Insurers tend to err on the side of more handoffs early on.
Same detection logic on voice, chat, WhatsApp, SMS, and email. Voice signal detection includes audio markers (tone, pace, breathing distress); text-channel detection works on language patterns and explicit content. Handoff happens immediately on voice; async channels route to the VC queue with the same context summary.
Every signal detection and every handoff is a Timeline event with severity, signal category, conversation excerpt, and handler routing. FOS files and Consumer Duty audits read directly from Timeline — no separate documentation project required.
30-day pilot. We configure detection against your existing vulnerable-customer signal library. You measure detection rate, handoff time, and handler outcomes. No platform migration. No long lock-in.
