Skip to main content
Demonstrates ArcKit’s approach to documenting AI-powered public services with stringent data protection and accessibility requirements.

​
Project Overview

A conversational AI assistant helping UK taxpayers navigate self-assessment, PAYE queries, and tax credits through natural language interactions in English and Welsh. View the complete project: arckit-test-project-v2-hmrc-chatbot

​
Key Features Documented

  • LLM architecture with PII redaction pipelines
  • Bilingual NLP handling (English/Welsh)
  • HMRC system integration patterns
  • Tax domain knowledge base design
  • Conversation audit and compliance logging
  • Accessibility (WCAG 2.1 AA) considerations

​
Generated Artifacts

The repository contains:
  • Conversational flow diagrams
  • AI model selection rationale
  • Data protection controls for tax data
  • Integration architecture with HMRC APIs
  • Fallback and escalation strategies
  • Performance and cost modelling
  • Responsible AI governance framework

​
Commands Used

# Initialise chatbot project
arckit init

# Core architecture design
arckit.design

# AI-specific design review
arckit.design-review --focus ai-ethics

# Generate compliance documentation
arckit.compliance --standards gdpr,wcag

# Cost and performance analysis
arckit.costs

​
When to Use This Pattern

Reference this example for:
  • Citizen-facing AI services
  • Chatbots handling sensitive personal data
  • Multi-lingual digital services
  • Tax or financial advice automation
  • Government conversational interfaces
  • Responsible AI implementation

​
Key Learnings

ArcKit documented a multi-layer approach: input sanitisation, prompt injection guards, output filtering, and audit logging. Critical for tax data.
Welsh Language Standards add architectural requirements. The design included language detection, model selection per language, and translation quality assurance.
Generated documentation covered model bias testing, explainability requirements, human oversight triggers, and continuous monitoring for drift.
Not all queries suit automation. ArcKit captured handoff logic to human advisors, preserving conversation context and ensuring seamless transitions.

Cabinet Office GenAI

Cross-government GenAI platform strategy

Scottish Courts GenAI

Justice sector GenAI with MLOps

Build docs developers (and LLMs) love

Get started for freeTalk to us