Toss Bank's AI-First Culture: 99.5% ID Forgery Detection, Faster Docs, Smarter Support

How Toss Bank Builds an AI-Driven Product Organization

Toss Bank is turning AI into an everyday utility across its products and operations. The goal: faster judgment, cleaner execution, and clear differentiation in customer experience. This isn't a side project-it's an organizational culture.

Org design: embed, centralize, and coordinate

AI isn't parked in a single team. Machine learning experts sit inside core business units like lending and deposits to solve field problems and ship features. Alongside that, specialized groups-the ML Platform Team, the Goddess ML Service Team, and the Product Development ML Service Team-support scale and reuse.

At the enterprise level, an AI task force (TFT) maps out new introduction areas, coordinates priorities, and sets the systems needed to spread AI across the company.

High-value use cases shipped to production

ID forgery and alteration detection runs on discriminant models trained on 100,000+ ID images, achieving 99.5% accuracy. This tightens risk controls without dragging down onboarding speed.

OCR turns financial documents into text at scale, covering more than 50,000 foreclosure documents each month. AI-based automation on top of OCR has pushed both speed and throughput in back-office workflows.

Support that scales with signal, not headcount

Digital accessibility in the mobile app helps customers find the right functions fast-mirroring the instant guidance you'd expect at a branch window. On voice, a homegrown speech-to-text (STT) model converts 150,000+ calls to real-time text every month, increasing operational efficiency and response speed.

Model strategy: general LLMs vs. domain precision

General-purpose LLMs struggle with financial context and Toss Bank's unique product names. To close that gap, the team is building its own language model with financial expertise baked in. Expected outcome: better counseling quality and more accurate responses.

Culture: AI for everyone, not just engineers

Employees who aren't developers are encouraged to use AI in their daily work. As usage spreads, AI becomes a shared way of working-beyond one-off tools or isolated pilots.

What product teams can borrow-today

• Place ML where value is created: embed experts in squads that own lending, onboarding, or servicing.
• Run a dual model: centralized platform teams for tooling and governance, embedded teams for outcomes.
• Pick use cases with clear ROI: ID verification, OCR-driven back office, voice analytics in support.
• Instrument everything: track accuracy, turnaround time, deflection, and error rates from day one.
• Close the loop: pipe model outputs back into product UX (guided flows, next-best-actions, proactive support).
• Own your domain: fine-tune or build models for product names, policies, and regulatory context.

Metrics that matter

• Forgery detection accuracy: 99.5% on 100,000+ ID images.
• Document throughput: 50,000+ financial documents processed monthly via OCR.
• Voice operations: 150,000+ calls transcribed to text monthly via STT.
• Time to resolution and deflection rates in support (tie metrics directly to product changes).

Implementation notes for regulated products

• Bias and drift checks: schedule evaluations by segment; monitor false positives/negatives on ID checks.
• Human-in-the-loop: set confidence thresholds that trigger manual review for high-risk actions.
• Traceability: log prompts, versions, features, and decisions for audits and incident response.
• Data minimization: keep only what you need; encrypt and control access by role.

If you're building identity flows, review industry guidance such as NIST's digital identity standards for assurance levels and risk controls. NIST SP 800-63 Digital Identity Guidelines

Next steps for your roadmap

• Week 0-2: Pick one use case with clear KPIs (e.g., ID check accuracy, document turnaround).
• Week 3-6: Ship a constrained POC with HITL and full logging; define go/no-go thresholds.
• Week 7-10: Integrate into the product flow; add user-facing guidance where AI adds certainty.
• Week 11-12: Run post-launch reviews; set model retraining cadence and drift alarms.

Want to upskill your product org on practical AI? Explore curated resources for jobs and teams. Complete AI Training: Courses by Job