TL;DR
AI voice agents do not replace contact centres — they replace the repetitive 60–80% of calls that don't need a human. The economics flip in favour of AI somewhere between 50 and 200 seats, depending on call mix, language coverage and AHT. The right question is not "AI or humans?" but "which calls, at what latency, at what cost per minute, with what fallback?"
Why this question matters now
Five years ago, AI voice was a novelty. Today, a well-deployed agent like WiseRep handles inbound tier-1 support, outbound collections, appointment booking and IVR replacement at human-equivalent CSAT — in Arabic and English, at sub-300ms latency, at a fraction of per-minute cost.
The board-level question is no longer "does AI voice work?" It is: "at our call volume, language mix and SLA, when does it actually pay back?" This post answers that with numbers, not slogans.
The cost stack of a traditional contact centre
Before you can model AI, you need an honest picture of what a human-only contact centre actually costs. Most CFOs underestimate this by 30–50% because they only count salaries.
| Cost line | Typical share |
| Agent salaries + benefits | 55–65% |
| Supervisors and QA | 8–12% |
| Real estate and workstations | 8–10% |
| Telephony and software licences | 5–8% |
| Training and attrition replacement | 8–12% |
| Compliance, recording, reporting | 3–5% |
A 200-seat MENA contact centre running 2-shift coverage typically lands between USD 3.5M and 5.5M per year, all-in. Cost-per-minute of handled call lands around USD 0.45–0.85, depending on language mix and location.
That is the number AI has to beat — not the salary line alone.
The cost stack of an AI voice agent
AI voice cost looks completely different. There are no salaries, no seats, no shift premiums — but there are inference, telephony and orchestration costs that scale with minutes.
| Cost line | Typical share |
| LLM inference (reasoning) | 30–40% |
| ASR (speech-to-text) | 15–20% |
| TTS (text-to-speech) | 15–20% |
| Telephony (SIP / carrier) | 15–20% |
| Orchestration + observability | 5–10% |
| Human fallback (10–30% of calls) | 10–25% |
For an Arabic + English deployment at production quality, all-in cost lands between USD 0.08 and 0.22 per minute of handled call — including the human fallback for the calls AI shouldn't take.
That is 3–6× cheaper than human-only at scale. But it is not free, and it is not constant — it scales linearly with minutes, where human centres have step-function costs.
When does AI voice actually pay back? The crossover model
The crossover depends on three numbers: call volume, automation rate, and language complexity. Here is the simplified model we use with clients:
Monthly savings = (Minutes × Automation rate × (Human cost/min − AI cost/min)) − Platform cost
Worked example for a 200-seat MENA bank running Arabic + English support:
- Monthly minutes: ~1.2M
- Automation rate (realistic): 65%
- Human cost/min: USD 0.65
- AI cost/min (all-in): USD 0.18
- Platform + integration amortised: USD 35,000/month
Monthly savings = (1,200,000 × 0.65 × 0.47) − 35,000 = USD 331,000/month
Annualised: ~USD 4M in net savings, with payback in 3–5 months for a typical WiseRep deployment.
Below ~50 seats the math gets thin — platform and integration cost dominates. Above ~100 seats, it almost always pays back within a year.
Latency: the budget that breaks most deployments
Cost is one constraint. The other is latency. Humans tolerate roughly 600–800ms of total response delay before a conversation feels broken. That budget has to cover:
| Stage | Budget |
| Telephony jitter + VAD | 50–100 ms |
| ASR (streaming) | 100–200 ms |
| LLM reasoning (first token) | 150–300 ms |
| TTS (first audio chunk) | 100–200 ms |
| Network egress | 50–100 ms |
| Total target | <700 ms |
Miss this and the agent feels robotic regardless of how good the reasoning is. Hit it consistently and even sceptical users stop noticing it isn't human.
The practical implications:
- Use streaming ASR and TTS — never batch.
- Choose regional model endpoints (UAE / EU) over global ones; saves 80–150ms.
- Prefer smaller, distilled LLMs for routing and intent; use large models only when reasoning depth is required.
- Pre-warm voices, cache greetings and common responses.
- Measure latency end-to-end in production, not just at the API boundary.
Where humans still win — and probably always will
AI voice is a force multiplier, not a replacement. There is a category of calls where you should not even try to automate, because the unit economics flip and the brand risk explodes:
- High-emotion calls — bereavement, complaints with regulatory exposure, safety incidents.
- High-value sales — anything above ~USD 5,000 of CLV where a human relationship pays back instantly.
- Highly ambiguous or multi-step disputes that require judgement, empathy and reading between lines.
- Calls that need legal commitment on behalf of the company.
- VIP and key-account handling where the relationship is the product.
The right design is AI-first, human-warm-transfer-always, not AI-or-nothing.
The 60/30/10 rule of thumb
After dozens of WiseRep deployments across banking, telecom, government and retail, the call-type distribution we see is remarkably consistent:
- 60% fully automatable — balance enquiries, status checks, password resets, appointment booking, FAQs, simple billing.
- 30% AI-assisted — AI handles intake, authentication and data capture; human handles the judgement call.
- 10% human-only — emotion, escalation, VIP, legal.
The payback math is built on the 60%. The 30% improves AHT and consistency. The 10% you leave alone — and you instrument carefully so AI never tries.
A practical decision framework
Use this checklist before committing to an AI voice rollout:
- Volume: are you above ~30,000 handled minutes/month? Below that, ROI is fragile.
- Call mix: what % of calls are repetitive, well-defined and low-emotion? If under 40%, savings shrink fast.
- Language coverage: do you need Arabic dialects + English code-switching? Budget for fine-tuning, not just licensing.
- Integrations: can the agent reach CRM, billing and identity systems in <300ms? If not, fix that first.
- Latency SLO: can your stack hold <700ms end-to-end? If you can't measure it, you can't ship it.
- Fallback: do you have a warm-transfer to a human agent within 2 seconds, with full context handed over?
- Governance: who owns the prompts, who reviews failures, who signs off on tone and compliance?
If you can answer all seven confidently, you are ready to deploy. If not, those gaps are the project plan.
What we measure on every WiseRep deployment
| Metric | Why it matters | Target |
| Containment rate | % of calls resolved without human | >65% on tier-1 |
| End-to-end latency (p95) | Conversation feel | <700 ms |
| CSAT vs human baseline | Customer-side quality | Within 5 pts |
| AHT reduction (assisted calls) | Operational impact | 20–40% |
| Cost per handled minute | Unit economics | <USD 0.22 |
| Escalation accuracy | Don't hand off cold | >95% with context |
If you are not tracking these from day one, you cannot prove ROI to the CFO when the renewal conversation comes around.
The honest answer to "will it replace my contact centre?"
No. It will change what your contact centre is for. The humans you keep will handle the 10–30% of calls where their judgement, empathy and authority matter most. The 60% that exhausted them and drove attrition will be handled at a fraction of the cost, in any language your customers speak, at any time of day.
That is a better contact centre. It is also a much cheaper one.
Take the next step
Thinking about AI voice for your contact centre? WiseRep is the Valus enterprise voice platform — Arabic and English, contact-centre scale, sub-300ms latency, deployed across banking, telecom and government in the MENA region. Book a working session:
Get in touchFAQ
At what scale does an AI voice agent pay back?
For most MENA enterprises, the crossover sits between 50 and 100 contact-centre seats. Below that, platform and integration cost dominates. Above 100 seats, an Arabic + English deployment typically pays back within 6–12 months.
What's a realistic automation rate?
60–70% containment on tier-1 inbound is a reasonable target for a well-scoped deployment. Pushing past 80% usually trades CSAT for cost, which is rarely worth it.
How is latency different from a normal API integration?
Voice is real-time. Every component — ASR, LLM, TTS, telephony — has to stream and meet a sub-700ms end-to-end budget. Batch APIs that work fine for chat will feel broken on a phone call.
Do AI voice agents work in Arabic?
Yes, but only with the right stack. MSA-only models collapse on Khaleeji, Egyptian and Levantine. Production-grade Arabic voice needs dialect-aware ASR, code-switched language understanding and natural-sounding TTS — all things WiseRep is built around.
What's the biggest mistake teams make?
Trying to automate the wrong calls. The economics work when you target the repetitive 60% — not when you try to replace human judgement on the hard 10%.



