A sovereign AI programme for a government ministry
Arabic-first models on national infrastructure, a governance framework a minister can sign, and citizen services moved onto it one at a time.
- Region
- Jordan
- Industry
- Government and public sector
- Engagement model
- Embedded Teams
- Duration
- Phased over a defined programme
- Client
- A government ministry
Context
A ministry in Jordan serves citizens who speak Levantine Arabic, read Modern Standard Arabic and switch to English for the words the state itself borrowed. Its services are delivered through counters, call centres and a portal built at different times by different vendors. The government's digital policy asks the ministry to move those services onto shared national infrastructure, to use artificial intelligence where it improves service, and to do so without citizen data leaving the country or decisions leaving the hands of officials.
The constraints are sovereignty and accountability. Jordan's Personal Data Protection Law governs how citizen data may be processed and where; national policy expects public-sector capability to be built, not rented; and procurement must be transparent enough to survive an audit and a change of minister. A model hosted abroad, trained on data the ministry cannot see and answering in an Arabic no Jordanian speaks is not a solution the ministry can sign.
The mandate is a programme, not a pilot: Arabic-first models deployed on national infrastructure, a governance framework a minister can sign, and citizen services moved onto the platform one at a time — the first by voice — with an embedded squad that transfers capability to civil servants as it works.
Mandate
The engagement is bought as Embedded Teams: a dedicated squad inside the ministry's digital directorate, led by an Altuon principal and accountable to the director, with a capability-transfer plan that names the civil-service roles that will take over and when.
- 01Deploy speech and language models that work in Levantine Arabic and Modern Standard Arabic, with English switching handled as normal, on the national cloud or the ministry's own infrastructure.
- 02Write the governance framework — data classification, residency, model register, human oversight, procurement and audit — in a form the minister signs and the audit bureau can inspect.
- 03Move the first citizen service onto the platform by voice, with a human official reachable from every step and every decision of consequence taken by an official.
- 04Add services one at a time through the same gates, each with its evaluation set and its sign-off.
- 05Keep every record, transcript, model and log inside the country, with the sub-processor list reduced to national providers where the framework requires it.
- 06Transfer the capability: civil servants trained and in post for each role in the squad before the squad steps down.
Approach
- 01
Discover
Map the ministry's services by volume, language and consequence; inventory the systems and the national infrastructure available; read the Personal Data Protection Law obligations and the national digital and AI policy with the ministry's legal advisers; establish which decisions officials must keep.
- 02
Define
Write the governance framework and the architecture: national data plane, model options that fit it, identity, audit and oversight; build the first Arabic evaluation set from recorded service interactions with consent and from synthetic dialect cases; agree the transfer plan naming roles and dates.
- 03
Build
Deploy models on national infrastructure; build the voice channel, the policy engine and the integrations to the first service's systems; build the model register and the audit trail; run the evaluation set on every change; train the first civil servants alongside the squad.
- 04
Prove
Open the first service to a defined district or caller segment with an official reviewing every interaction; report accuracy, resolution, handover and residency compliance to the steering committee; widen as the numbers hold; obtain the sign-offs the framework requires.
- 05
Operate
Add services through the same gates; run the monthly review; step the squad down role by role as civil servants take over, with overlap for each; leave the framework, the register and the runbooks with the directorate.
What was built
- Governance framework
- Classification, residency, model register, oversight, procurement and audit rules, written for signature by the minister and inspection by the audit bureau.
- Arabic-first models on national infrastructure
- Speech and language models for Levantine and Modern Standard Arabic with English switching, deployed inside the national data plane.
- Voice channel
- Telephony and dialect-aware speech for the first citizen service, with recording notice and consent handled under the Personal Data Protection Law.
- Policy engine and official handover
- A rules layer deciding what the system may do and what an official must decide, with handover carrying the transcript and the citizen's request.
- Service integrations
- Adapters to the first service's records and case systems, then to each service added, through a governed integration layer.
- Model register and audit trail
- Every model's owner, purpose, evaluation results, versions and review date; every interaction and decision recorded and attributable.
- Evaluation harness
- Arabic evaluation sets per service and dialect, adversarial cases, and the gate that blocks a release that fails.
- Capability-transfer plan and training
- Named civil-service roles, training paths, overlap periods and the record of each handover.
Outcome
A blueprint measures what the mandate promised. The figures below are the ones this programme would report to the steering committee and the audit bureau; they are filled from the record when a client approves publication, and are otherwise shown as the measures rather than as numbers.
- Services moved onto the sovereign platform
- Measured against the baseline agreed in the mandate
- Arabic response accuracy on the evaluation set
- Measured against the baseline agreed in the mandate
- Data residency exceptions at go-live
- Measured against the baseline agreed in the mandate
Timeline
Stack
Bring us a mandate like this one.
A proposal request takes seven short steps and is read by the engagement lead who would run the work. References for real engagements are provided under non-disclosure, matched to your industry and region.