Voice AI Concurrent Calls: How to Size Capacity

How many concurrent calls does your voice AI need? A step-by-step sizing method using Erlang workload, peak intervals, campaign headroom and load tests.

Voysera teamUpdated 2 min read

Voice AI Concurrent Calls: How to Size Capacity

To size voice AI concurrent calls, multiply the calls in your busiest interval by the average call length and divide by the interval length. That gives the average number of simultaneous calls, called the workload in Erlangs. Then add headroom for random bursts and campaigns, and prove the number with a load test. "Unlimited scale" is not a sizing plan.

What is the formula?

Workload in Erlangs = calls in the interval × average handle time in seconds ÷ interval length in seconds. It comes from the queueing work of A. K. Erlang, first published in 1917 (Assembled). Size each 15 or 30 minute interval separately, because averaging a whole day understaffs the peaks, which is where service is lost (LogicCommons).

A worked example

  • Busiest hour: 1,200 calls.
  • Average handle time: 180 seconds.
  • Workload: 1,200 × 180 ÷ 3,600 = 60 concurrent calls on average.
  • Random arrivals make the count swing around that average. With 60 on average, about 95 percent of moments stay at or below roughly 73 simultaneous calls (60 plus 1.645 times the square root of 60).
  • Add any outbound campaign running at the same time, and keep a separate reserve for inbound.

How is AI different from sizing human agents?

Human staffing must add shrinkage for breaks, training and absence, about 30 percent by industry average (Assembled). AI agents have no shrinkage and answer instantly, so there is no queue to plan around. Instead, every simultaneous call needs real capacity in four places, and the weakest one sets your limit.

BottleneckSymptom when it runs outWhat to check
Speech recognitionSlow or missing transcriptsConcurrent stream limits
Language modelLong pauses before repliesTokens per minute and requests per minute quotas
Voice synthesisChoppy or delayed speechConcurrent synthesis streams
TelephonyBusy signals or failed callsSIP trunk channels and carrier limits

How do you size without guessing, step by step?

  1. Pull 3 to 6 months of call data and find the busiest 30 minute intervals.
  2. Calculate the workload for each with the formula above.
  3. Add burst headroom, about 95th percentile rather than the average.
  4. Add planned outbound campaigns, with a reserve kept for inbound.
  5. Check each bottleneck in the table against that number.
  6. Load test on a production-like stack and confirm replies stay under 800 milliseconds at peak (Hamming AI).

How does Voysera handle capacity?

Capacity is reserved for each customer and sized to their volume, with an inbound reserve so campaigns never crowd out service lines. Voysera scales horizontally in our cloud or on your own servers, and runs across several AI providers and regions with automatic fallback, so one slow provider does not take your line down.

Watch live load against capacity on a voice AI call center dashboard, and plan campaign volume with AI outbound calls in Arabic. Request sizing guidance for your traffic.


المكالمات المتزامنة في الذكاء الاصطناعي الصوتي: كيف تحدد السعة

لتحديد عدد المكالمات المتزامنة، اضرب عدد المكالمات في أكثر فترة ازدحامًا في متوسط مدة المكالمة، واقسم على طول الفترة. الناتج هو متوسط المكالمات المتزامنة (الحمل بوحدة إرلانغ). ثم أضف هامشًا للذروات العشوائية والحملات، وأثبت الرقم باختبار حمل.

مثال

  • أكثر ساعة ازدحامًا: 1,200 مكالمة، ومتوسط المدة 180 ثانية.
  • الحمل: 1,200 × 180 ÷ 3,600 = 60 مكالمة متزامنة في المتوسط.
  • في 95٪ من اللحظات يبقى العدد عند 73 مكالمة أو أقل تقريبًا.

نقاط الاختناق الأربع

  • التعرف على الكلام: حدود البث المتزامن.
  • النموذج اللغوي: حصص الرموز والطلبات في الدقيقة.
  • توليد الصوت: البث المتزامن.
  • الهاتف: قنوات SIP وحدود المشغل.

في ڤويسيرا السعة محجوزة لكل عميل مع احتياطي للمكالمات الواردة، وتعدد المزودين والمناطق مع تحويل تلقائي.

تحدث إلى المبيعات.