A Gulf tax authority - Arabic NLP, agent assistance

The problem

Employees and service agents needed precise answers from extensive Arabic tax guidance, without slow manual searches and without relying on unsupported general-model knowledge. An agent on a live contact has minutes, and tax guidance is long, written in formal Arabic, and unforgiving of approximation. Both available options were bad: search several guides while the taxpayer waits, or defer the question and generate a follow-up contact that would not have existed otherwise.

What we built

We implemented a retrieval-augmented question-answering pipeline over the Gulf tax authority's documents - extracting and chunking source content, retrieving relevant passages with multilingual embeddings, and prompting JAIS and other Arabic-capable models to produce grounded responses inside an agent-assist interface.

What it turned on

The service became faster not by trusting the model more, but by reducing the evidence an agent had to inspect and making the boundary between sourced guidance and model generation explicit - retrieval relevance, factual support, Arabic language quality, and output format were evaluated separately, and weak-evidence cases withheld a definitive draft rather than producing one.