Re-architecting a federal ministry's flagship citizen-services platform.

At a glance.

All client identifying details are anonymised under NDA. A senior lead reference call is available to qualified prospective clients on request.

The ministry operates one of the federation's most visible citizen-services platforms. It handles a high-millions-per-year volume of transactions across roughly forty service categories, in Arabic and English, against strict service-level commitments published in the federation's annual government-performance report. The platform had been built on a stack assembled over a decade and was carrying material technical debt: response times had degraded, the Arabic conversational experience was visibly weaker than the English one, and the call-centre escalation rate had been climbing for six consecutive quarters.

The ministry had two non-negotiable constraints. First, no service interruption during the cut-over: the platform is referenced by name in the federation's published service charter and a visible outage would have been politically untenable. Second, full in-country data residency on a federally-approved sovereign cloud, with every model — including the conversational layer — running inside the client perimeter. Off-the-shelf conversational products were ruled out for that reason alone.

We embedded an eighteen-person pod for nine months, reporting to the ministry's permanent secretary. The pod was structured into three streams running in parallel: platform re-architecture, Arabic-first conversational layer, and cut-over engineering. The cut-over plan was the spine of the engagement: every architectural decision was tested against the requirement that any individual citizen-facing service could be switched to the new platform independently and reverted within a single working hour. The Arabic conversational layer was built on the firm's in-house NLU stack, evaluated against a corpus the ministry's communications team curated, and tuned in weekly working sessions with the call-centre supervisors who own the escalation queue. First production cut-over of a single service category happened in week eleven; full migration completed at month nine.

Seventy-three percent of citizen interactions are now resolved without human escalation, against a baseline of forty-one percent on the legacy platform. Median response time across Arabic and English is four-point-six seconds. The call-centre escalation queue has been reduced sufficiently that the ministry has reallocated headcount from level-one support into proactive citizen outreach. Zero service interruption during the cut-over. The platform is now referenced internally as the federation's reference implementation for Arabic-first government services and has been visited by two other federal entities for transfer.

Three. First, in citizen services the conversational layer's quality in Arabic is not a feature, it is the product — and that quality is unattainable on off-the-shelf English-first stacks. Second, cut-over engineering deserves a dedicated stream from week one; it is the artefact that makes the ministry's risk position tenable. Third, embedding the pod with the call-centre supervisors — not with the IT function — was decisive in calibrating the escalation logic.