AI for Government · Citizen services · Correspondence · Internal departments
Answer every citizen, in Arabic. Finish the paperwork.
- Saudi Arabic, day and night
- In the Kingdom or in your own building
- Identity before any file is read
- Officers keep every decision
04 · WHERE THE WORK LANDS
Citizen services
→ Status explained
logged · recording kept
Correspondence
→ Letter routed
clock started
Permits
→ Appointment booked
logged · recording kept
Complaints
→ Case escalated
inside the clock
Departments
→ Circular cited
source attached
Field teams
→ Visit scheduled
report drafted
03 · HARNESS
VOHO GOVERNMENT AI HARNESS
ANSWER · VERIFY · COMPLETE · ACCOUNT
VOICE
Listen · speak · barge-in
TOOLS
Case · archive · calendar
MODELS
Arabic speech · LLM · OCR
AGENTS
One per service
IDENTITY
Before any file is read
ESCALATION
Decisions go to officers
EVALUATIONS
Scored on real requests
MODEL ROUTING · CONTEXT · RAG · GUARDRAILS · PII REDACTION · HUMAN DECISIONS · AUDIT TRAIL
02 · SYSTEMS OF RECORD
01 · WHERE CITIZENS REACH YOU
Swipe the drawing sideways.
One AI layer. Every channel a citizen uses and every system behind it.
The Voho Government AI Harness
A model that cannot see the case file is a leaflet.
It does not know where the request is, what is missing, or what the procedure says this month, and it must never be the thing that decides a citizen’s eligibility. The harness gives models the systems, the identity checks and the escalation rules of a government entity, and records everything they do inside them.
A model on its own
- Recites a procedure that changed last year
- Cannot see whether the request moved
- No idea who is on the phone
- Promises a callback nobody receives
- Leaves nothing an auditor can follow
The same model, in the harness
- Reads the live case file once identity passes
- Quotes the current circular, with the clause attached
- Books the appointment in the real calendar
- Sends the upload link and restarts the clock
- Writes the whole interaction onto the case
Systems of record
What your entity already runs
- Case management
- Correspondence system
- Nafath
- Government service bus
- SAP / ERP
- CRM
- GIS
- Document archives
- Payments
- IVR / contact centre
- SMS gateway
Voho Government AI Harness
The runtime between your systems and the models
- Tool calling
- Model routing
- Government connectors
- Identity verification
- Memory
- Context
- RAG
- Workflow orchestration
- Permissions
- Escalation rules
- Human decisions
- Evaluations
- Guardrails
- PII redaction
- Audit trails
Model layer
Routed per task, never one model for everything
- Saudi Arabic speech
- Speech synthesis
- Foundation models / LLMs
- Arabic document and handwriting models
- Classification models
- Your own models
Agent layer
Scoped to one service each
- Citizen Services Agent
- Correspondence Agent
- Permits Agent
- Complaints Agent
- Inspections Agent
- Internal Department Agent
Where the work lands
Finished in the system that owns it
- Requests answered
- Letters routed
- Appointments booked
- Complaints escalated
- Documents drafted
- Inspections scheduled
01
Answer.
Every call, message and portal question picked up in Saudi Arabic or English, at any hour, on the numbers and channels citizens already use.
02
Verify.
Identity checked against your own rules, through Nafath where your entity is integrated. Nothing about a file is read out before that passes.
03
Complete.
The request moved in the system that owns it: status explained, document requested, appointment booked, letter routed to the right section.
04
Account.
Every interaction, source and action recorded, so the answer a citizen was given can be reconstructed months later, by you or by the regulator.
01 · Citizen services
The call at nine at night, answered and finished.
Most citizen calls are not complicated, they are just unanswered: where is my request, what is missing, when can I come in. Here is one of them with the lid off, including the line where the agent stops and leaves the decision to an officer.
Transcript
Identity
Nothing on the file is readable until verification passes.
Tool calls
waiting…
Policy
- Telling a citizen what their own file says, and booking them a slot, is inside the delegated scope.
- An exemption, an eligibility call or anything discretionary is drafted with the rule attached and left for an officer.
Illustrative call · recording, transcript and every tool call written to the case
02 · Correspondence and archives
Arabic paper, read and routed before lunch.
Letters still arrive on paper, in Arabic, sometimes handwritten, often with the attachment missing. They are read, checked against the checklist, routed to the section that owns them and answered in draft, with the uncertain line flagged rather than guessed.
01 · Arrives
طلب ترخيص نشاط تجاري
Commercial activity licence request
شكوى تأخر في إصدار شهادة
Complaint: certificate delayed
تعميم من الوزارة بخصوص المواعيد
Ministry circular on appointment hours
طلب إعفاء من رسوم
Fee exemption request
استفسار عن حالة معاملة
Enquiry on a transaction status
طلب زيارة تفتيشية
Inspection visit request
02 · Read and checked
applicant_name
محمد بن سعد العتيبي
national_id
1XXXXXXXXX
subject
طلب إعفاء من رسوم
reference
IN-2026-04417
received_date
1447-03-14 · 2026-09-16
handwritten_note
مرفق ما يثبت الحالة
The handwritten line reads at 0.71 and is marked for a person rather than guessed.
Identity document attached
Signed application form
Supporting evidence for the exemption
Subject matches an existing case file
03 · Routed and drafted
Section
Licensing · Fees and exemptions
Matched on subject and on the three previous letters from this applicant.
Priority
Standard
Clock
4 of 5 days left
Drafted reply · awaiting an officer
نشكر لكم تواصلكم. طلبكم قيد الدراسة، ويلزم إرفاق ما يثبت الحالة لاستكمال النظر فيه.
We have your request. It is under review and needs the supporting evidence attached before it can proceed.
Nothing is sent, and no exemption is decided, without an officer
Across the entity
One harness. Every counter and every section.
The telephony, the connectors, the identity rules and the evaluation set are built once. The next service is an agent and a workflow on the same foundation.
Citizen services
The queue that never empties
- Request and application status explained against the actual case file
- Requirements and fees answered from your published procedures, with the clause cited
- Appointments booked and rescheduled into the service centre calendar
- Complaints logged, classified and escalated inside the regulatory clock
- Callbacks and reminders made automatically when a file is waiting on a citizen
Correspondence and archives
Arabic paper, at volume
- Incoming letters and forms read in Arabic, including handwriting and stamps
- Classified, summarised and routed to the section that owns them
- Attachments checked against the checklist before a file moves
- Decades of archived correspondence answerable in one question, with the page cited
- Draft replies prepared in the entity’s own letter format for an officer to sign
Internal departments
The work behind the counter
- HR, procurement and finance assistants scoped to what each employee may see
- Policies, circulars and regulations answered with the paragraph attached
- Inspection visits scheduled and field reports written up from dictation
- Tenders and contracts read and compared against the template
- Handover notes and meeting minutes drafted in Arabic
Connectors
One AI layer. Every system behind the counter.
Nothing is replaced to get the first service live. Where there is an API we use it, where there is a database we read it, and where there is neither we work through the screens your staff already use.
Case management
Requests, statuses, service levels
Correspondence
Incoming and outgoing, with attachments
Nafath
Identity, where your entity is integrated
Service bus
The integrations you already expose
SAP / ERP
HR, procurement, finance
Archives
Scanned correspondence and decisions
Contact centre
Cisco, Avaya, SIP. Your numbers
SMS and payments
Notifications, links, fee references
Built for a public entity
What the committee asks, answered in the product.
Residency, disclosure, decision rights and retention are configuration your own teams set and can audit, not promises in a proposal.
Inside the Kingdom, or inside your building
Run it in a Saudi region or on your own servers, including entities with no outbound internet. Citizen data never leaves the boundary your policy sets.
Never trained on citizen data
What a citizen says or uploads is used to serve that request. It is not used to train models, ours or anyone else’s.
Officers keep the decisions
Eligibility, exceptions, fee waivers and anything discretionary are drafted with the rule attached and left for a named officer. The agent never decides them.
Identity before disclosure
Nothing on a file is read out until the caller passes your own verification. Failed checks end the call or go to a person, and are logged either way.
Answers you can defend
Procedural answers cite the circular and the clause they came from, so a citizen can be shown why, and an auditor can see where it came from.
Built for your obligations
Retention windows, PII redaction in transcripts, per-role access and a full audit trail, configured by your own teams to support your PDPL duties.
From the library
What this would look like at companies you know.
Demonstrations, not customer work. We built five agents for each of these organisations from public knowledge of how they run; none of them asked, and none of them run this today. Open one to see this page’s argument applied to an entity you recognise.
Elm Company
Government
The company behind much of the Kingdom’s digital government, where a service outage is a citizen who cannot complete something they are legally required to do.
Run the demos →Zakat, Tax and Customs Authority
Government
Tax, zakat and customs administration, where a taxpayer’s question is usually about a rule that has a precise answer and a deadline attached.
Run the demos →General Organization for Social Insurance
Government
Social insurance, where a caller is usually asking whether a benefit they are counting on will actually arrive.
Run the demos →Saudi Data and AI Authority
Government
The national data and AI regulator and builder, where the question is usually whether a proposed use of data is permitted.
Run the demos →Monsha’at
Government
The small and medium enterprise authority, where a founder calling about funding usually has a different problem from the one they are describing.
Run the demos →How a project runs
Start with one service. Keep the harness.
The first service proves itself on a number you already report: answered calls, average handling, letters cleared within the clock.
01
Take one service
One licence, one permit, one benefit. Something with real call volume, a published procedure and a status your case system already holds.
02
Connect inside your network
Read access to the case system, the procedure library and your telephony, reviewed by your information security team before a call is routed.
03
Replay real requests
The agents are scored against calls and letters that already happened: what they would have answered, what they would have routed, where they would have escalated.
04
Go live out of hours first
Evenings and weekends, where the alternative is an unanswered phone, then widen. The next service reuses the same harness.
What entities ask before anything is connected.
If your question is not here, bring it to a technical session with our engineers.
01Does it make decisions about citizens?
No. It explains what a file says, moves a request along your existing procedure, books appointments and drafts replies. Eligibility, exceptions and anything discretionary are prepared with the rule attached and left for an officer to decide.
02Where does the data live?
In a Saudi region or on your own servers, including entities with no outbound connection. The models run where the data is, and recordings and transcripts stay inside the boundary your policy sets.
03Can it verify who is calling?
Through your own rules, and through Nafath where your entity is already integrated. Until verification passes, the agent has nothing about the file. What each action requires is yours to set.
04Our correspondence is Arabic, and some of it is handwritten.
That is the normal case here, not the hard one. Letters, forms and stamps are read in Arabic including handwriting, with confidence scores on every field, and anything the model is unsure of is flagged for a person rather than guessed.
05Our systems are old and were not built for this.
Most are. Where there is an API we use it, where there is a database we read it, and where there is neither we work through the screens your staff use. Nothing is replaced to get the first service live.
06What happens when it cannot help?
It hands over to your officer with the summary and everything already done, or books the citizen a slot rather than leaving them in a queue. The call is never dropped into a dead end.
07How do we know it is giving citizens correct answers?
Procedural answers are generated from your own published procedures with the clause cited, not from the model’s memory, and the agents are evaluated against real past requests before go-live and monitored after it.