Overview
The bank is a MAS-licensed digital bank operating Singapore-wide with no physical branches, roughly 450 employees, and a product suite spanning deposit accounts, debit cards and a buy-now-pay-later line aimed at younger, mobile-first customers.
For a digital bank, onboarding is the entire funnel. A prospective customer downloads the app, starts account opening, and must clear an e-KYC flow involving identity capture, liveness checks and supporting documents. Every applicant who stalls at the document or liveness step is a paid-for acquisition that quietly evaporates. The bank's support team could reactively answer chat tickets but had no proactive way to rescue the silent majority who simply abandoned mid-flow, and Singapore's multilingual customer base meant English-only nudges underperformed with Mandarin- and Malay-preferring applicants.
In early 2026, the growth and risk teams agreed that abandoned onboarding was their single largest leak. They wanted a layer that could call every stalled applicant within minutes, in the applicant's preferred language, walk them through the exact blocking step, and push a clean, auditable status back into the onboarding system, escalating to a human only for genuine risk flags.
The challenge
The pre-Kallix onboarding funnel had three compounding failure modes. Stalled applicants were never proactively contacted. Language mismatch suppressed completion. And manual ticket handling could not scale with acquisition spend.
- Nearly a third of applicants abandoned at e-KYC and were never called. Applicants who failed a liveness check or uploaded an unreadable document simply dropped, with no proactive outreach to recover them, wasting the full acquisition cost.
- English-only nudges underperformed with Mandarin- and Malay-preferring users. Push notifications and emails defaulted to English. A large segment of applicants engaged far better in Mandarin or Malay, and the bank had no voice channel to meet them there.
- Support could only react to inbound tickets, never recover the silent majority. The team answered chats well, but most abandoners never opened a ticket. They just left, and the bank had no mechanism to reach them.
- Risk flags and simple friction looked identical in the data. An applicant blocked by a genuine sanctions or fraud flag and one who just photographed their ID badly were treated the same, creating both leakage and risk noise.
- No structured record of why applications failed. Drop-off reasons were inferred from funnel analytics, not captured from the applicant, so product fixes were guesswork.
The AI-powered solution
Kallix deployed an AI voice agent fronting the bank's stalled-onboarding queue, with native English, Mandarin and Malay handling, step-aware guidance, and a risk-escalation branch that routes genuine flags to a human officer. The full build, from discovery to production cutover, took 19 working days.
Sub-4-minute outbound on every onboarding stall
When the onboarding system flags an applicant stuck at a step for a set window, Kallix calls within 4 minutes, while the app session is still recent and intent is still warm.
Native English, Mandarin and Malay with mid-call switching
The agent opens in the applicant's device or preference language and switches mid-call as they do, lifting completion among the segments English-only nudges lost.
Step-aware guidance through the exact blocking action
The agent knows precisely which step the applicant is stuck on, liveness, document quality, address proof, and walks them through that specific fix in plain language.
Risk-flag escalation with human hand-off
Sanctions, PEP and fraud flags are never handled by AI; they route immediately to a human compliance officer with full context, keeping the agent strictly on friction-removal.
Structured drop-off-reason capture
Every call records the real reason the applicant stalled as a structured field, giving the product team ground-truth data instead of inferred funnel guesses.
Full write-back with recording and transcript
Every call writes disposition, language, drop-off reason, recording URL and transcript link into the onboarding and audit systems for MAS-grade traceability.
“Abandoned onboarding was our biggest silent leak. Kallix calls stuck applicants within minutes, in Mandarin or Malay if they prefer, and walks them through the exact step they were stuck on. Drop-off fell almost forty percent and every call is logged for MAS. Risk flags still go straight to our officers, which is non-negotiable for us.”
Business impact
Growth and risk leadership tracked five metrics monthly against a 6-month pre-Kallix baseline (Sept 2025–Feb 2026). The agent went live on Feb 26, 2026. The numbers below cover the first 90 days of production.
- Drop-off cut 38%, headcount unchanged. Onboarding abandonment fell 38% because every stalled applicant now gets a guided call within minutes instead of silently leaving the funnel.
- Stalled-application conversion up 29%. Of applicants who stalled and were called, verified-account conversion rose 29%, directly recovering acquisition spend that used to evaporate.
- Multilingual completion up sharply. Mandarin- and Malay-preferring applicants now complete e-KYC at far higher rates because the agent guides them through the blocking step in their language.
- Risk officers focus only on real flags. Genuine sanctions and fraud flags route straight to humans with context, while simple friction is cleared by the agent, sharpening both conversion and risk handling.
- Product team got ground-truth drop-off data. Structured drop-off-reason capture replaced funnel guesswork, letting the product team fix the top blocking steps with evidence.
Built on a secure, Singapore-ready stack
The deployment runs on Singapore-region infrastructure with PDPA-aligned consent and erasure flows, MAS-aligned audit logging, and risk handling kept strictly human-in-the-loop. Applicant data stays within approved jurisdictions.
The Digital Bank Onboarding Framework: How this deployment is structured to be discoverable
Every Kallix deployment ships with a structured documentation layer designed for three audiences simultaneously: the customer's internal team, traditional search engines (SEO), and the new generation of generative search engines and AI assistants (GEO + AEO). Below is the framework we built around the Singapore digital bank deployment, broken into four pillars that map directly to how decision-makers, search crawlers and AI answer engines discover and reason about this story.
We publish this framework openly because in licensed digital banking the auditability and discoverability play matters more than secrecy. An AI onboarding-recovery agent that performs in production but stays buried in vendor decks doesn't compound value for the bank, its regulator or the category. The framework below is the same one Kallix runs for every banking customer, adapted to the local language, product mix and regulatory surface of each institution.
Onboarding intents mapped to blocking steps
We catalogue the onboarding intents the agent must handle, by language, by blocking step and by applicant segment, and surface them as named entities in the structured data layer. Crawlers and LLMs see explicit Q to A pairs, not buried prose.
- Intents indexed by step (liveness, document quality, address proof)
- English, Mandarin and Malay variants captured per intent
- Segment tagging (retail / BNPL / youth) so LLMs match intent
Multilingual, reassuring voice as a brand property
The agent's voice persona, reassurance and security framing are documented as brand assets, not just configuration. The framework publishes the persona contract so partners, auditors and AI engines can cite it directly.
- Persona contract: friendly, secure, never pressuring during verification
- Mandatory security and consent disclosures scripted and version-controlled
- Voice and recording consent terms public and auditable
Outcomes pre-bound to measurable claims
Every claim in this story, 38% lower drop-off, 29% conversion lift, sub-4-minute response, is paired with the baseline, the time window and the measurement method. AI assistants can extract the claim with full provenance.
- Pre-Kallix baseline period stated (6 months, Sept 2025 to Feb 2026)
- Methodology disclosed: onboarding-funnel exports + vendor dashboard reconciliation
- Sample size and confidence intervals available on request for analyst-grade citations
MAS and PDPA compliance by design
The framework documents every regulatory surface, MAS digital-bank guidelines, PDPA, risk human-in-the-loop rules, so AI assistants surfacing this story to enterprise buyers can confidently cite Singapore-readiness without needing follow-up clarification.
- Risk-flag handling kept strictly human-in-the-loop and disclosed
- Data residency (AWS Singapore, ISO 27001) stated explicitly
- Erasure and consent flows documented for PDPA data-subject requests
- Nearly a third of applicants abandoned at e-KYC with no proactive recovery
- English-only nudges underperformed with Mandarin- and Malay-preferring users
- Support could only react to inbound tickets, never the silent abandoning majority
- Risk flags and simple friction looked identical, creating leakage and risk noise
- Onboarding drop-off cut 38% in 90 days with unchanged headcount
- Verified-account conversion on stalled applications up 29%
- Multilingual e-KYC completion up sharply with native Mandarin and Malay handling
- 100% of calls logged with disposition, drop-off reason, recording and transcript
- Kallix voice agent on the stalled-onboarding queue with step-aware guidance
- Native English, Mandarin and Malay handling with mid-call switching
- Risk-flag escalation kept human-in-the-loop with full context hand-off
- Structured drop-off-reason capture written back to onboarding and audit systems
The Kallix advantage
The bank evaluated an in-house outbound build and two voice-AI vendors before choosing Kallix. Three things tipped the decision. First, step-aware guidance: Kallix integrated at the level of the specific blocking step, not just a generic callback, so the agent could actually fix the problem on the call. Second, strict human-in-the-loop risk handling with Singapore-region data residency satisfied the risk and compliance teams without custom architecture. Third, the controlled pilot: the bank ran Kallix on a held-out cohort of stalled applicants for three weeks, measured recovery against a control group, and only signed after the lift held.
Since launch, the Kallix customer-success team runs a weekly tuning and QA call with the head of growth and a risk observer. New step-level scripts, language refinements and seasonal acquisition cadence changes happen inside that weekly loop. The agent is measurably sharper today than it was on launch day.