Engineering Specification for Enterprise Voice AI Agents
Specification Matrix
Core Engineering Requirements
- ›An enterprise voice AI RFP must demand P95 latency under 280ms measured from caller silence to first received audio packet.
- ›Dialect evaluation must test code-switching: colloquial Moroccan Darija intermixed with French commercial terms.
- ›Never procure a voice agent without deterministic state-machine guardrails to prevent unapproved commitments or price hallucinations.
- ›Demand direct ERP integration (REST, ODBC, or SOAP) with rollback capabilities in case of transactional failure.
1. The Four Non-Negotiable Procurement Pillars
When issuing a Request for Proposal (RFP) for enterprise voice AI, Moroccan CIOs must reject superficial vendor demonstrations and evaluate candidates across four rigorous engineering gates: (1) Turnaround Latency under realistic carrier conditions (P95 < 280ms), (2) Code-switched Moroccan Darija transcription accuracy (WER < 8.5%), (3) Deterministic state management preventing commercial hallucination, and (4) Sovereign on-premise or sovereign VPC hosting compliant with CNDP Law 09-08.
2. SLA, Error Budget & Verification Matrix
Require vendors to contractually commit to verifiable metrics: a 99.95% uptime availability SLA (unplanned downtime under 21.6 minutes/month), maximum concurrent call drop rates below 0.1%, acoustic barge-in interruption reaction under 65ms, and automated failover routing to human supervisor queues whenever conversational confidence drops below 85%.
| Engineering Metric | Minimum Threshold | Target Production SLA | Breach Consequence |
|---|---|---|---|
| End-to-End Turnaround Latency (P95) | < 320 ms | < 280 ms | 10% billing credit per 50ms excess |
| Call Availability Uptime | 99.90% | 99.95% | Automated SLA tier penalty rebate |
| Barge-In Audio Silence Cutoff | < 90 ms | < 55 ms | Mandatory DSP audio loop refactor |
| Darija Code-Switch Word Error Rate | < 12.0% | < 8.5% | Vendor re-training at zero client cost |
3. Fallback & Graceful Degradation Architecture
Production voice AI systems must anticipate connectivity hiccups, model timeouts, and unhandled dialect queries without disconnecting the customer. Vext mandates a 3-tier fallback hierarchy: Tier 1 executes rapid local intent matching via compiled regex and finite state machines; Tier 2 queries high-speed local inference kernels; Tier 3 triggers an instantaneous SIP REFER (RFC 3515) warm-handoff to human agents, transmitting conversational transcript JSON directly to the agent CRM screen before audio bridging.
// Standard Vext Telephony Handover Payload (Sent to Agent CRM via WebSocket)
{
"call_id": "sip-cas-2026-0314-88412",
"caller_number": "+212661000000",
"detected_locale": "ar-MA-Darija",
"intent_history": ["order_status_inquiry", "carrier_delay_complaint"],
"sentiment_score": -0.68,
"last_transcript": "Bghit n3ref fin wsel l-colis dyali dial Casablanca",
"recommended_action": "escalate_to_logistics_tier2",
"erp_order_ref": "CMD-2026-8941"
}4. Production Security & Real-Time PII Scrubbing
Under no circumstances should raw credit card numbers, national identity identifiers (CIN / CNE), or bank account numbers (RIB) be stored in plain text logs or passed unmasked to third-party inference models. Production voice architectures must incorporate streaming Named Entity Recognition (NER) and regex masks directly within the speech recognition pipeline, replacing sensitive tokens with cryptographic pseudonyms before transcript persistence.
Specification FAQ
Need technical guidance on this specification?
Speak directly with our systems engineers in Casablanca.