Metro Bank · AI Voice Biometrics · 2019

AI Voice
Biometrics

Designed the end-to-end experience for AI-powered voice authentication in Metro Bank's telephony channel, replacing forgotten credentials with conversational identity verification, reducing authentication time by 81% and saving £1.2M annually.

Role

Head of Product Design

Timeline

2019

Cross-functional

CX, Operations, Fraud, Compliance

Customer Impact

40K+ calls/week

81%
Reduction in auth time
76%
Customer enrolment rate
£1.2M
Annual saving

Forgotten
credentials.
Frustrated
customers.

Metro Bank's call centre handled 40,000+ calls per week. Our data revealed that 38% of callers could not complete traditional credential-based verification on their first attempt. Customers needed their Customer ID, password, and security number, but with only 1–2 logins per week, these were frequently forgotten.

Even fallback methods had a 22% failure rate. The result: frustrated customers locked out of their accounts, call handling times averaging 6+ minutes, and a measurable impact on NPS. AI-powered voice biometrics was brought to the table as a potential solution. I took ownership of evaluating the concept, got aligned on the direction, and led the full experience design, machine learning that could build a unique voiceprint from natural conversation and passively authenticate the customer without them knowing it was happening.

The design challenge was enormous: trust and transparency (consent to voice recording without perception of surveillance), regulatory compliance (FCA, GDPR), and a layered fallback architecture that didn't rely solely on a single authentication factor.

Authentication Pipeline, Voice Biometrics System
FCA Approved
🔒 GDPR Compliant
🛡️ Anti-Spoofing
Enrolment Flow (First-time Setup)
Step 1
Customer opts in, gives consent via app
12s avg
Step 2
3–5 training calls recorded, voiceprint created
~2-4 calls
Step 3
Voice model stored securely, testing begins
Ready in 24h
Authentication & Matching Algorithm
Audio Capture
Extract voice during call
t+0–20s
Feature Extraction
MFCCs, spectral patterns
t+20–45ms
Model Matching
Compare to voiceprint
t+45–120ms
Decision Output
Confidence score (%)
t+120–150ms
Confidence Thresholds & Actions
95%+ confidence: Immediate authentication, proceed
85–94% confidence: Request one-time PIN for extra verification
75–84% confidence: Fallback to security questions + phone verification
<75% confidence: Manual agent review, retry with full PIN
Edge Cases Handled
• Noisy environments (noise filtering)
• Cold, hoarse, emotional voices
• Accents & speech variations
• Aging effects on voice (retraining)
• Multiple calls to build confidence
Fraud & Anti-Spoofing
• Liveness detection (breathing, speech patterns)
• Speech synthesis detection
• Replay attack prevention
• Anomaly scoring per call
• Fallback to additional verification
76%
Enrolment Rate
23s
Auth Time (vs 120s)
98.7%
Match Accuracy
40K+
Auth Calls/Week
£1.2M
Annual Saving

Consent.
Confidence.
Fallback.

I owned the end-to-end experience design, working directly with the Director of CX, Head of Customer Operations, and Fraud & Compliance leadership. The approach balanced invisible AI with visible trust.

01

Customer journey mapping, Mapped every telephony authentication pathway, identifying 14 distinct failure points and their root causes across the call centre operation.

02

Transparent enrolment experience, Designed the opt-in flow for voice biometrics. Key insight from testing: customers were significantly more willing to enrol when framed as "your voice is your password" rather than "we will record your voice."

03

Passive voice sampling architecture, Designed the implementation so sampling initiated immediately when a representative picked up. No script changes, no awkward prompts. The AI analysed natural conversation in the first 15–20 seconds.

04

2FA safety net, Even with 95%+ confidence voice match, the system prompted for one lightweight confirmation (last 4 digits of phone, or recent transaction amount), ensuring no single-factor reliance.

05

60+ moderated research sessions, Tested enrolment, passive authentication, and edge cases (background noise, illness, third-party callers) across call simulations to validate customer trust and accuracy.

"Customers described the experience as 'effortless' and 'like they already know me.' Removing authentication friction didn't just save time, it changed the relationship between customer and bank."

Effortless
authentication.
Measurable
savings.

Voice biometrics transformed telephony from Metro Bank's biggest friction point into a competitive advantage, with enrolled customers reporting the highest NPS scores across all banking channels.

Authentication Speed

81% reduction in average auth time, from 97 seconds to 18 seconds for enrolled customers.

Adoption

76% of eligible customers enrolled within 6 months, exceeding the 30% target by 2.5x.

Resolution Rate

92% first-call resolution for voice-authenticated customers, up from 61%.

Cost Savings

£1.2M estimated annual saving in reduced call handling time and failed authentication rework.

Next Case Study

Almosafer:
MENA's First AI Travel Experience

Want in?

Score 100% on the Design Eye test to earn access. Or enter the password if you have it.

4 rounds. No partial credit.