Pillar III: Ethics, Transparency, Interpretability · § 04

Inclusivity and accessibility

Bizzi is used by chief accountants with twenty years of experience, by interns in their first month, by users with visual impairments, and by clerks in provincial offices on slow connections. The AI works for all of them. Accessibility is not a feature flag added late. It is a constraint shaping the UI from the first wireframe.

Context

The accountants who use Bizzi rarely fit one persona. SMEs have small accounting teams who rely on AI to compensate for headcount. Enterprises have larger teams with stricter audit requirements. A confidence indicator which conveys meaning only through colour fails users with colour-vision deficiency. An HITL approval flow which needs a mouse fails users on keyboard-only workflows. An LLM which ignores Vietnamese diacritics in retrieval fails most of our customers.

Accessibility. WCAG 2.2 AA

Contrast. Text-on-background ratio at least 4.5:1 for normal text and 3:1 for large text.
Colour is not the only channel. Confidence is rendered as colour and icon and text. A red-green confused user still sees the warning.
Keyboard-only. Every HITL approve, edit, and reject is reachable without a mouse, with visible focus state.
Screen readers. Every AI component carries ARIA labels. The confidence indicator is narrated in words, not only colour.
Focus management. When the AI opens an HITL modal, focus moves into the modal. Escape closes it and returns focus to the trigger.
Timeout warning. Five minutes before a session expires, the user is warned and given a one-click extension.

Language support

Bizzi operates primarily in Vietnamese with English available where customers need it:

UI in Vietnamese, fully translated.
AI output in Vietnamese for every user-facing interaction (chatbot, summary, suggestion).
AI input multilingual. OCR and the LLM handle Vietnamese, English, and bilingual invoices.

For Vietnamese specifically, OCR is fine-tuned on a diacritic-complete corpus. Vector embeddings normalize diacritics so a search for “hoa don” still finds “hóa đơn”. Confidence thresholds are calibrated so faded diacritics do not unfairly penalize an otherwise clean extraction.

For enterprise customers with expatriate staff, we provide an EN/VI toggle stored in user preferences, bilingual document export for audit, and English-capable customer support.

Connectivity and device

Vietnam has wide variance in internet quality between major cities and provincial offices. The UI is designed for both:

Bandwidth-aware. AI output loads progressively. The page does not block on the LLM response.
Offline-resilient. The HITL queue and recent history are reviewable offline against already-synced data.
Mobile-first. Many accountants triage HITL on a phone between meetings. The mobile UI is not an afterthought of the desktop one.

Inclusivity in our process

Inclusivity holds only if we test it. We run user research with cohorts which vary by age, experience, region, and ability. An accessibility audit by an external specialist runs once a year. Pilots with SME customers happen before any broad rollout. SME accounting teams rely on AI more heavily than enterprise teams, and they surface usability gaps the larger customers do not.