Generative AI in Vietnamese banking 2026: from POC to production

By 2026, most of Vietnam's top-15 banks have moved Generative AI from POC to production. This article distills the use cases that actually return ROI, the architectures being chosen, and the compliance friction under Decree 13 and the SBV's new AI circular.

1. Four GenAI use cases delivering real ROI in Vietnamese BFSI

• Internal copilot for RMs and contact-center agents. RAG over thousands of internal docs for product, fee and process lookup. AHT down 25–35%, new-hire onboarding 40% shorter.
• SME credit-file summarisation. LLM reads financials, statements and contracts → drafts the credit memo for the analyst. Per-file processing drops from 4 hours to 45 minutes.
• KYC data extraction and document-fraud detection. Vision-LLM reads CCCD, business licences and passports, cross-checks against internal data, flags retouched documents and selfie deepfakes.
• Personalised letters and segment-level marketing copy. Generates per-segment email/SMS copy, lifting conversion 12–20% over static templates.

2. The winning architecture: governed RAG, not large-model fine-tuning

By mid-2025 many banks had tried fine-tuning Llama-70B or Qwen-72B on internal data — and most stopped. Reasons: retraining cost after every product change is high, and the audit trail of a fine-tuned model is hard to defend to the SBV. The settling architecture is: foundation model (GPT-5/Gemini 2.5/self-hosted Qwen-32B) + versioned RAG + output guardrails + full prompt/response logging for traceability.

3. Compliance: Decree 13, customer data and the onshore boundary

Decree 13/2023 requires sensitive personal data to be processed with a clear legal basis, with a DPIA, and — in many cases — kept onshore. This creates a hard boundary: prompts containing PII cannot be sent to off-shore LLMs without tokenisation/encryption. The common 2026 pattern is a two-tier split: large off-shore models see only tokenised data; smaller self-hosted models (Qwen, Mistral) onshore handle anything containing PII.

4. Realistic ROI and a recommended investment frame

A mid-sized Vietnamese bank rolling out three internal use cases over 9 months typically sees 8–15 bn VND CAPEX and 1.5–2.5 bn VND/year OPEX. Typical payback is 14–20 months from contact-center, underwriting and marketing savings. Customer-facing use cases (chatbot, voicebot) pay back more slowly but compound retention.

5. Five recommendations for Vietnamese BFSI CIOs in 2026

• Start with internal use cases — fast ROI, low compliance exposure.
• Don't lock in a single LLM; use an LLM-gateway pattern so you can swap models.
• Invest in RAG pipeline and data hygiene before investing in GPUs.
• Ship output guardrails and run periodic evals — don't deploy and forget.
• Design for AI audit from day one, not as a retrofit.