ZTABS builds AI voice agents for Riyadh enterprises in government citizen services, banking, healthcare and enterprise customer service. Riyadh concentrates government ministries, corporate headquarters and major healthcare facilities with significant call volumes across Arabic and English. AI voice agents that handle Gulf Arabic naturally address caller expectations for spoken interaction rather than rigid IVR menus. ZTABS is remote-first by design from our Houston, Texas office.
Riyadh voice agent demand spans government citizen services and unified hotlines, banking customer service and account management, healthcare scheduling and triage routing, and enterprise contact center automation.
AI voice agent development in Riyadh at a glance
| Need | What we build | Key requirement |
|---|
| Government | Citizen services, hotline automation | Arabic dialect, PDPL |
| Banking | Account servicing, card management | SAMA sector rules |
| Healthcare | Scheduling, triage routing, intake | Health data scope |
| Telecommunications | Customer service, plan management | CST, bilingual |
| Enterprise | Contact center automation | Scalability, multilingual |
How voice agents work
A voice agent receives or places calls through telephony infrastructure. Incoming audio converted to text by speech-to-text. Text processed by language model for intent and response. Response converted to speech by text-to-speech.
STT selection. Provider and model evaluated for actual language, dialect, noise and channel. Riyadh callers include Gulf Arabic (Najdi dialect), MSA, English and South Asian language speakers. Gulf Arabic differs from MSA in pronunciation, vocabulary and grammar. STT models trained primarily on MSA may not handle dialectal speech accurately without evaluation or tuning.
TTS selection. Provider and model for naturalness, latency and language. Arabic voice persona evaluated for natural Gulf Arabic prosody. English voice for bilingual callers. Pronunciation of government service names, banking terms and medical terminology verified.
Model selection. Language model for the specific task. Government service routing needs different model behavior than banking account inquiries. Arabic-English code-switching handling evaluated per actual caller patterns. Model size balanced against latency.
Building a voice agent for Riyadh government or banking?
Tell us about your call volume, use case and compliance requirements. We start with assessment covering telephony infrastructure, model evaluation and regulatory mapping.
Inbound versus outbound
Inbound voice agents answer calls from citizens, customers or patients. Risk-approved identity verification and step-up authentication before revealing protected information or making changes. Handle requests within the owner-approved action envelope. Transfer to human when exceeding authorised scope.
Outbound voice agents place calls for appointment reminders, service notifications or follow-ups. Saudi direct-marketing rules require consent for personal commercial communications and an effective stop mechanism; actual scope depends on the entity type and call activity. Service notifications distinct from marketing. Recording consent disclosed at call start. CST (Communications, Space and Technology Commission) service-provider obligations assessed per actual entity registration and activity.
Conversational design
Identity and verification. Before revealing account details, health information or service status, verify caller identity through risk-approved methods appropriate to the data sensitivity. Static identifiers alone are not adequate authentication for consequential actions. Match verification level to action risk, enforce tenant-level and object-level authorisation, and do not collect sensitive credentials into audio, transcripts or logs.
Arabic dialect handling. Gulf Arabic callers in Riyadh use Najdi dialect vocabulary and expressions that differ from MSA. STT models evaluated for dialectal speech. The voice agent aims to understand natural spoken Arabic as Riyadh callers actually speak, including code-switching between Arabic and English for brand names, technical terms and product names.
Speech endpointing. Detecting when caller finished speaking versus pausing. Arabic conversational patterns may involve longer pauses. Configurable per context.
Interruption handling. Barge-in detection. Agent stops and processes. Essential for experienced callers who know what they need.
Repair and confirmation. Low-confidence triggers confirmation. Read-back of dates, account numbers, service reference numbers. Explicit confirmation before consequential actions. Model confidence alone does not authorise expanded actions.
AI identity and recording disclosure. The voice agent identifies itself as an automated system at call start and offers a human alternative, as an agreed design choice separate from recording consent. Recording disclosure and consent follow immediately.
Latency. Response time varies by the STT, model and TTS provider chain. The project measures median, P95 and P99 end-to-end latency on agreed dialects and call scenarios during evaluation, not as a fixed guaranteed response time.
Saudi recording and communications
Recording and interception of communications in Saudi Arabia is subject to applicable law. The agreed consent-first architecture discloses recording at call start and obtains caller acknowledgment as a design choice; this is distinct from determining the actual lawful basis for each processing activity. Callers who decline recording cannot have audio capture or live STT silently continue; architecture provides a lawful alternate path such as unrecorded human transfer or manual service. Even when stored audio recording is off, live STT transmission and transcript storage remain privacy processing. Actual recording obligations and lawful bases assessed per entity type and sector regulation.
Safe bounded workflow
Low confidence. Transfer to human when intent unclear rather than guessing.
Unsupported intents. Acknowledge, explain limitation, offer human transfer.
Refusal and opt-out. Human transfer through staffed queue or accessible manual alternative during agreed service hours. No forced bot interaction. Callers outside staffed hours offered voicemail, callback where supported or self-service per business rules.
Caller distress. Medical emergency or acute distress: tested escalation path to appropriate resources. Urgent callers can use independent emergency services without waiting for bot callback or queue. Voice agent is not emergency service, clinician, lawyer or regulated adviser and does not guarantee immediate connection to any external service.
Failed transfer. Callback offered where staffing and authorisation support it. Not an unqualified promise of callback for every failed transfer.
After-hours. Voicemail, callback scheduling where supported, emergency information or limited self-service per business rules.
Unknown outcome. When API, payment or scheduling completion is uncertain, reconcile the current remote object state and authoritative local ledger before retry. Logging a failure alone is insufficient. Bounded retries with limits and manual exception path. Receiver-side business-effect deduplication and atomic local booking or entitlement updates, distinct from outbound provider-endpoint idempotency. No exactly-once guarantee; current object and slot check before committing a booking to avoid race conditions.
Government voice agents
Riyadh concentrates Saudi government ministries and entities with significant citizen call volumes.
Citizen service hotlines. Government entities operate unified numbers for citizen inquiries. Voice agents handle routine service information, application status checks, document requirement inquiries and appointment scheduling where the entity provides API access. Integration with government platforms requires the entity's approved programme and confirmed permissions.
Application status. Citizens call to check application progress. The voice agent queries the entity's system through approved APIs with tenant-level authorisation and provides status information. No autonomous disclosure of information beyond what the entity has authorised for the specific caller verification level.
Appointment scheduling. Booking appointments at government service centers. Integration with the entity's scheduling system through confirmed APIs. Booking confirmed only after authoritative system acknowledgment.
Multilingual support. Arabic as primary language with English for expatriate callers. Language detection at call start routes to the appropriate language pipeline. South Asian language support assessed per actual caller demographics and available provider coverage.
Banking voice agents
Riyadh banking institutions handle significant customer call volumes for account servicing.
Account servicing. Balance inquiries, recent transactions, card status, payment processing. Risk-approved identity verification before revealing account details. Integration with core banking through the institution's approved APIs with tenant and object-level authorisation. SAMA sector rules assessed per actual institution type and service scope.
Card management. Card blocking, limit changes, replacement requests, PIN inquiries. Multi-factor identity verification for consequential card actions. Changes require explicit confirmation.
Transfer and payment. Wire transfer status, standing order management, bill payment. Transaction authorisation within agreed limits and verification levels. Transfers above agreed thresholds escalated to human agents or step-up authentication.
Loan and financing inquiries. Information about existing financing products, payment schedules, early settlement inquiries. No autonomous lending decisions or product recommendations. Advisory elements transfer to qualified staff.
Healthcare voice agents
Riyadh healthcare facilities handle scheduling, triage routing and patient inquiries.
Appointment scheduling. Patients call to schedule, reschedule or cancel. Integration with hospital scheduling systems through provider-approved APIs with actual tenant role and object-level authorisation. Scheduling writes require the provider's approved programme, customer-vendor agreement and confirmed write permissions. Booking confirmed only after authoritative system acknowledgment.
Triage routing. Clinical symptom-to-care-level routing requires actual intended-use assessment, accountable clinical authority, approved protocol, tested escalation and human review boundaries; scheduling and intake are nonclinical functions handled separately. No autonomous clinical decisions. No credentials in audio; secure alternate entry for sensitive identifiers.
Patient intake. Pre-visit demographics, insurance information and chief complaint. Health data subject to additional PDPL Article 23 controls, not excluded from the law. Encryption, minimum necessary access with workforce role controls and audit logging.
Prescription refill. Risk-approved identity verification, medication confirmation, pharmacy selection, refill submission through approved systems with confirmed permissions. No sensitive identifiers spoken aloud; secure alternate entry where authentication requires protected data.
Integration requirements
Government systems. Integration through entity-approved APIs with confirmed access and write permissions. Authentication and data scope per the entity's approved programme.
Banking systems. Core banking, CRM through institution-approved APIs. Tenant role and object-level authorisation. Write permissions per the institution's approved programme.
Healthcare systems. Hospital information systems through provider-approved APIs. Write permissions require the provider's approved programme and customer-vendor agreement.
Telephony. SIP trunking, numbers, concurrent capacity. Provider selection for regional latency.
Voice-specific cost drivers
Voice agent development follows the custom software cost guide as general platform engineering context. Actual costs depend on selected provider usage, telephony, STT, TTS, LLM, concurrency, minutes and retention scope.
| Cost driver | Description |
|---|
| Telephony infrastructure | SIP trunking, numbers, concurrent capacity |
| STT provider usage | Per-minute pricing by provider, model and language |
| TTS provider usage | Per-character or per-minute by voice and language |
| Model inference | Per-token pricing, scales with call volume |
| Concurrency | Simultaneous call capacity licensing |
| Recording and storage | Consent-gated configured coverage, transcript storage |
| Monitoring and evaluation | Quality monitoring, accuracy tracking |
Cost and timeline
Voice agent development as general custom platform engineering uses four guide tiers.
| Scope | Cost range | Timeline |
|---|
| Internal Tool | $25K to $60K | 6 to 14 weeks |
| Business Application | $60K to $150K | 14 to 28 weeks |
| Enterprise Platform | $150K to $300K | 28 to 40 weeks |
| Large-Scale System | $300K to $500K+ | 40 to 60+ weeks |
2026 US vendor benchmark
Ranges are 2026 US-buyer guide-based planning benchmarks for a mid-complexity business application: multi-role access, 3 to 5 integrations, custom workflows and a modern web frontend.
| Vendor type | Typical cost | Timeline |
|---|
| Freelancer or dev collective | $20K to $75K | 4 to 10 months |
| Offshore agency | $40K to $120K | 5 to 11 months |
| Nearshore agency | $70K to $200K | 4 to 8 months |
| US agency (ZTABS tier) | $90K to $280K | 4 to 8 months |
| Big 4 or systems integrator | $250K to $800K+ | 8 to 16 months |
Ranges are 2026 US-buyer guide-based planning benchmarks for the scoped business-application comparison. Where applicable to the agreed engagement, legacy-system integration (SAP, Oracle or mainframe interfaces) and compliance engineering or readiness preparation (SOC 2, HIPAA safeguards or PCI DSS controls) carry an additional guide allowance of $20K to $80K across vendor paths. Independent CPA SOC 2 examination and opinion are scoped and priced separately from internal readiness engineering. The guide uses 15 to 25 percent of build cost annually as an ongoing-maintenance planning estimate; actual support scope and fees are agreed separately. These are scoped planning allowances, not an empirical market survey or voice-package rate card.
Proposal evaluation criteria
- Relevant voice workflow experience and customer-approved action authority with supporting evidence
- Engineering and integration reliability including unknown-outcome recovery, bounded retries and manual exception paths
- Assessed privacy, security and regulatory boundaries with clear independent assurance responsibilities
- Speech, language and accessibility evaluation with held-out test sets and agreed acceptance criteria
- Transparent scope, milestones, change control, operating fees and ongoing support terms
- Human handover, training, validation and support with actual customer and provider responsibility allocation, agreed staff and user languages, and documentation and translation responsibilities
Need a scoped estimate for Riyadh voice agent development?
Share your use case, call volume and compliance requirements. We provide a scoped estimate based on system complexity after reviewing both four-tier scope and vendor benchmark context.
Saudi regulation and data protection
Saudi PDPL. The Personal Data Protection Law applies to personal data processing within the Kingdom and, where applicable, to processing of Saudi residents' data outside the Kingdom. SDAIA is the competent authority. Lawful bases, purpose limitation, data subject rights, breach notification and cross-border transfer conditions apply. Article 23 imposes additional controls on health data. Article 24 imposes additional controls on credit data. These are supplementary sector obligations, not blanket exclusions from the law. Assessment per actual entity role and data scope.
SAMA banking. The Saudi Central Bank regulates banking and financial institutions. Customer communication standards and data handling assessed per actual institution type and service scope. Credit data subject to PDPL Article 24 additional controls.
Healthcare regulation. Health data subject to PDPL Article 23 additional controls alongside the general PDPL framework. Voice agents processing patient data comply with applicable health data requirements. Clinical symptom-to-care-level routing requires actual intended-use assessment, accountable clinical authority, approved protocol, tested escalation and human review boundaries; scheduling and intake are nonclinical functions handled separately.
Recording and communications. Recording consent disclosed at call start as part of a chosen consent-first architecture. Architecture provides lawful alternate path for declining callers. Live STT transmission and transcript storage are privacy processing subject to assessment regardless of whether stored audio recording is active.
CST. The Communications, Space and Technology Commission (formerly CITC) regulates telecommunications services. Service-provider obligations, commercial communication rules and opt-out mechanisms assessed per actual entity registration and activity scope.
Evaluation methodology
Representative test calls. Expected uses, edge cases, accents, noise, channel quality on agreed test scenarios.
Held-out dialects. Riyadh linguistic diversity: Najdi Arabic, MSA, English, South Asian languages. STT accuracy across representative categories on the agreed test set.
Task completion. Intent correctly identified and action completed. Measured against human baseline.
False action and false refusal. Tracked separately. Government service false actions have different risk than retail false actions.
Adversarial prompts. Injection, social engineering, authorisation bypass.
Privacy leakage. Agent does not reveal other callers' data or internal systems.
Illustrative Example
Hypothetical: Riyadh government entity citizen services voice agent
A Riyadh government entity receives thousands of calls weekly for service inquiries, application status checks and appointment scheduling. The entity operates in Arabic and English with significant call volumes during peak service hours.
A voice agent would answer inbound calls, detect language (Gulf Arabic or English), verify caller identity through risk-approved methods, query the entity's systems through approved APIs with confirmed access permissions and tenant-level authorisation, provide service information and application status, and schedule appointments with authoritative system confirmation. Recording consent disclosed at call start.
Assumptions: entity API access approved through the entity's programme with confirmed permissions, service information documented with designated authority, recording consent architecture designed with lawful alternate path for declining callers, and telephony supports SIP. Residual risks include STT accuracy with Gulf Arabic dialects and code-switching, API rate limits during peak hours, complex multi-service routing edge cases, and caller preference for human agents. No guarantee of specific automation rate, wait time reduction or caller satisfaction outcome.
This is a hypothetical scenario for illustration. Actual scope, timeline and cost depend on your requirements.
How delivery works
ZTABS is remote-first by design from our Houston, Texas office. Houston is Central time; Riyadh is Arabia Standard Time (UTC+3), 9 hours ahead during Central Standard Time and 8 hours ahead during Central Daylight Time. Houston morning overlaps with Riyadh evening for coordination calls and handoffs.
When ZTABS is not the right fit
- Simple IVR tree with no AI. Traditional IVR vendors handle press-1 menus.
- On-site embedded team for this engagement. ZTABS is remote-first by design. On-site staffing requires separate arrangement.
- Arabic dialect speech model research for this engagement. Voice agent delivery uses evaluated STT providers; fundamental Arabic speech research is a separately scoped engagement.
- Clinical decision support for this engagement. Routing to clinicians per approved protocol, not autonomous clinical decisions.
- Regulatory compliance consulting for this engagement. Compliance engineering is part of platform delivery; independent compliance advice is a separately agreed engagement with qualified advisers.
White-label AI products
ZTABS builds custom white-label AI products for clients to brand and resell. Riyadh enterprises can commission white-label voice AI platforms for distribution.
Monitoring and operations
Real-time dashboards. Call volume, completion rates, transfer rates, STT accuracy, latency percentiles. Per-department and per-language metrics.
Model updates. Tested in staging. A/B testing. Rollback capability. Healthcare and government model updates reviewed by designated authority.
Cost monitoring. Per-call costs across STT, TTS, model inference and telephony. Monthly reporting by department.
Recording and transcription. Actual permitted configured coverage per consent architecture, not automatic every-call full capture. Failures, minimisation, access controls, retention, deletion, legal hold and approved lawful test-data reuse specified per engagement. No automatic cross-customer data reuse.
Telephony architecture
SIP trunking. Direct connections for quality and latency. Redundant trunks for failover. Local numbers provisioned per use case.
Concurrent capacity. Peak planning with headroom. Government services peak during working hours. Banking and healthcare have predictable daily patterns. Capacity planning sized per agreed peak forecasts, not infinite automatic scaling without infrastructure commitment.
Call routing. By caller history, department, language preference and time of day. Priority routing for urgent cases. Overflow when capacity exceeded.
Enterprise and telecommunications voice agents
Riyadh enterprises and telecommunications operators handle significant customer service volumes.
Telecommunications. Subscriber account inquiries, plan management, billing, device support. Integration with operator BSS/CRM through approved APIs. High-volume routine inquiries handled by voice agent with complex issues transferred to human agents.
Enterprise contact centers. General customer service, order management, appointment scheduling. Multi-channel integration. Seasonal volume management.
Handover
Voice agent handover includes runbooks covering telephony configuration, model versions and provider accounts, monitoring dashboards with agreed metrics, evaluation methodology with representative test sets, the agreed Saudi consent architecture and documented escalation procedures with staffed-queue contacts.
Ready to build a Riyadh voice agent?
Share your call volume, use case and compliance requirements. We start with assessment covering telephony, model evaluation and regulatory mapping.