Pillar III: Ethics, Transparency, Interpretability · § 05

Cultural norms and sensitivity

AI does not operate in a cultural vacuum. A model trained mostly on English-language data misreads Vietnamese business reality. It ignores Tết. It misses non-standard fiscal years. It uses a tone which lands wrong with finance teams. Bizzi builds for the Vietnamese market first and adapts to international contexts second.

Context

The mismatches which come from cultural assumptions are quiet but expensive. An anomaly detector which flags every salary bonus paid the week before Tết generates noise which erodes user trust in days. A chatbot which addresses a chief accountant as “bạn ơi” loses the account before lunch. Locality is not flavour text. It is correctness.

Tết (Lunar New Year)

Widespread closures. Most Vietnamese businesses close for 5 to 10 days around Tết. The AI does not flag missing approvals during the window.
Pre-Tết cashflow. Spending spikes before Tết. Payroll, bonuses, partner gifts. The anomaly detector treats this as seasonal pattern, not unusual behaviour.
Q4 invoice surge. Most Q4 invoices arrive in December. Capacity scales on its own in the final week of the Gregorian year.

Anomaly baselines are aligned to the lunar calendar week-by-week. Cashflow forecasting uses a Tết-aware model. The platform’s auto-scaling watches the calendar, not only the load curve.

Non-standard fiscal year

Many Vietnamese businesses run a fiscal year which does not align with January to December, often because of holding company alignment or industry regulation. The AI does not assume January is the start of the fiscal year. Each tenant configures its fiscal period, and reporting calendars follow it.

Currency and numbers

VND is the default. Large amounts are commonly abbreviated “tr” (triệu), “tỷ”, and “k” in correspondence. The LLM parses all of them.
Number formats. Both 1.000.000 and 1,000,000 parse to the same value. So do 1tr, 1tỷ, and 1m in context.
Foreign currency invoices. The AI asks for the exchange rate at booking time rather than guessing a daily mid-market rate.

Language and tone

Formal but warm. The chatbot uses formal Vietnamese without slipping into “bạn ơi” or stilted “kính thưa quý khách”.
No emoji in formal communication. Status messages and notifications do not use emoji unless the user has enabled a casual mode.
Respect address conventions. “Anh” and “Chị” are the default forms of address. The AI does not address users by first name without invitation.
Disambiguate abbreviations. “KH” means either khách hàng (customer) or kế hoạch (plan). The AI resolves from context and surfaces the resolved interpretation.

Tax authority interactions

The tax authority is the source of truth for tax IDs and e-invoice approvals. The AI does not silently correct a tax ID even when it looks malformed. If the invoice does not match the tax authority record, the transaction flags for human review instead of auto-rejecting.

Hard-coded guardrails

There are categories the AI does not generate output for, regardless of prompt:

Personal tax advice (Bizzi is a B2B product).
Personal investment or financial advice.
Legal judgement on a vendor or contract. The AI extracts data. It does not adjudicate.
Competitive assessments of the customer’s own competitors.

When we expand internationally

For multi-country tenants, the defaults become configurable. Fiscal year. Language. Currency. Address conventions. Locale-aware tests run in CI for every UI change. The cultural assumptions baked into the AI are documented in the Model Card so international customers review them before going live.