Regulatory AI · Fertilizer & Clean Ammonia (Enterprise)
UAE VAT Advisory Agent for Fertiglobe
FTA-reference, amendment-aware tax guidance grounded on curated FTA sources, deployed to Microsoft Teams, at 98% accuracy across 100+ eval cases.
- Role
- AI Engineer
- Designed the agent, grounding strategy, response contract and the escalation topic with context hand off.
- Industry
- Fertilizer & Clean Ammonia (Enterprise)
- Timeline
- Delivered - v1
Business impact
- accuracy
- 98%
- evaluation cases
- 100+
- citation
- FTA article reference
- deployment
- Microsoft Teams
Overview
FTA-reference, amendment-aware tax-guidance agent grounded on curated Federal Tax Authority source documents for verifiable, citable answers.
Deployed to Microsoft Teams for day-to-day finance use, reaching 98% accuracy across 100+ evaluation cases with every answer carrying its governing reference.
Technical architecture
- Conversational Surface
- Copilot Studio agent with curated prompt starters and a legal disclaimer
- Knowledge
- The Decree Law and amendments prepared into a searchable knowledge source
- Grounding Controls
- General knowledge off and web off, so the agent cannot answer outside the statute
- Response Contract
- Direct answer, Article citation, bullets only for genuine lists
- Escalation
- System escalate topic into a customer service queue with four structured context variables
Key features
- Closed world grounding on the Decree Law and its three amendments, with general knowledge and web access off
- Article level citation on every answer - a determination without its Article is not acceptable output
- Deliberate terseness - one or two sentences with no unrequested related articles or amendment history
- Scenario determination returning the VAT treatment plus the governing Article
- Statutory terminology used with the law's own defined terms and capitalisation
- Explicit gap handling where the answer requires regulation outside the knowledge base
- Human escalation carrying a question summary, relevant Articles, topic area and history to a live advisor
- High content moderation with a scoped fallback and a confidentiality disclaimer on feedback
- Curated prompt starters across the highest traffic VAT areas
Challenges solved
- Fabricated Article numbers are the failure mode that matters - mitigated by closing the world entirely
- Models over answer legal questions, so the instruction set spends real effort on what not to say
- Recognising the edge of the knowledge base and declaring it rather than approximating across it
- Separating informational guidance from advice, with formal opinions and disputes routed to a human
Tech stack
- Microsoft Copilot Studio
- Azure AI Search
- Azure OpenAI
- Dynamics 365 Omnichannel
- RAG
- Structured Instruction Engineering