AI delivery
Arabic conversational AI for a UAE government digital service
A UAE government digital services entity delivering high volume citizen and resident services online.
- Industry
- Government and Public Sector
- Pillar
- AI delivery
- Client geography
- United Arab Emirates
- Delivery location
- Dubai and Abu Dhabi
- Languages
- Arabic C2, English C1
- Engagement model
- Fixed scope implementation followed by managed operation
- Timeline
- Twenty-two weeks to public launch
The challenge
The entity's existing chatbot answered a narrow set of questions in stilted formal Arabic and failed on anything phrased naturally. Deflection was low, user satisfaction was lower, and citizens were routing back to the contact centre after failed self-service attempts, which made the channel a net cost. The entity was also clear that an AI assistant giving a wrong answer about a legal entitlement or a required document was an unacceptable outcome, which ruled out an unconstrained generative approach.
What Corpshore did
We designed a retrieval-grounded assistant rather than a free generation system. Every answer is grounded in the entity's own published procedural content, with citation back to the source page, and the assistant is explicitly constrained from answering outside its grounded corpus. When it does not know, it says so and hands off cleanly.
The Arabic work was the substance of the engagement. We built out the query understanding layer to handle Gulf dialect phrasing, code-switching between Arabic and English, common misspellings and the difference between colloquial and formal ways of asking the same procedural question. Our Arabic language team rewrote the underlying procedural corpus into plain, correct, consistent Arabic before any of it was indexed, because a retrieval system grounded in unclear source material returns unclear answers.
Human review sits over the whole operation. Sampled conversations are reviewed daily against accuracy and tone criteria, and failures feed back into the corpus and the retrieval configuration weekly.
Delivery model
Fixed scope implementation over twenty-two weeks, followed by a managed operation covering monitoring, review, corpus maintenance and continuous improvement. All processing configured to meet the entity's data residency and protection requirements.
Results
- Self-service containment rose substantially against the previous assistant, measured as sessions resolved without human handoff.
- Escalation to the contact centre from the digital channel fell correspondingly.
- Arabic query understanding accuracy improved markedly against the entity's test set, with the largest gains on dialect-phrased and code-switched queries, which the previous system had handled worst.
- Zero substantiated incidents of the assistant providing incorrect procedural guidance across the first operational period, which was the entity's primary acceptance condition.
Why it worked
The constraint was the design. Grounding every answer and refusing to answer outside the corpus produced lower coverage than an unconstrained model would have claimed, and it produced an assistant a government entity could actually put in front of citizens.
Discuss an engagement like this
All UAE enquiries answered within six hours.
This is a representative engagement. The client is anonymised, and the figures are drawn from engagement reporting. We do not name a client without written consent, and we do not publish a number we cannot evidence.