How to Design a Voice AI Surge Plan for Contact Centers

A voice AI surge plan is a written answer to three questions: how many calls your AI agents can hold at once during your busiest hour, what tells your team that limit is getting close, and where callers go once it is reached. Most voice AI evaluations never ask them. A demo shows that the agent handles one call well on an ordinary day. A surge plan shows that the deployment still works when an outage, a billing run, or a seasonal peak sends several times your normal volume in a few hours.

This guide is for contact center, IT, and operations leads who want that answer before signing, not after the first bad week.

Why does a normal-week evaluation hide surge risk?

Because missed calls do not grow in step with call volume. While arrivals stay under capacity, nearly every call is answered and the miss rate looks small and steady. Once arrivals pass capacity, every extra call waits or is lost, so misses climb much faster than volume does.

Here is an illustration, not a measurement. A line takes 5,000 calls in a normal week and misses 1%, so 50 callers go unanswered. Then a surge week brings 20,000 calls with the same capacity. If 15% are missed, that is 3,000 callers, and many of them call back and add to the load. Nothing in the normal week warned you, because the weekly average never came close to the limit.

Surges arrive in hours, not spread across a week. So evaluate in 15 or 30 minute intervals, and compare the two cases directly:

For the staffing side of peaks (shift design, self-service, deflection), see our guide to scaling customer support.

How many concurrent calls do you need at peak?

Start with this: peak calls per hour × average call length in minutes ÷ 60. The result is the number of calls active at the same moment, called concurrency, which is the unit voice AI capacity is sold and limited in.

Say your busiest hour brings 1,200 calls averaging 4 minutes. That is 1,200 × 4 ÷ 60 = 80 calls active at once, on average, across the hour. Add headroom, because calls bunch up inside the hour during a surge. Then add any outbound campaigns that draw from the same capacity.

How this works on Retell:

• Concurrency is the number of simultaneous active calls. If 15 callers are talking to your agents at the same time, that is 15 concurrent calls.
• Limits apply per workspace. Inbound, outbound, web, and batch calls in a workspace all draw from one pool.
• One agent can take many calls. Each agent can handle an unlimited number of calls as long as the workspace stays within its limit, so you do not need copies of an agent to cover a surge.
• There is a sizing calculator. On the Settings > Limits page, enter calls per busy hour, average durations, and pickup rate, and it recommends a concurrency limit, an inbound reservation, and calls per second (CPS), each with headroom for spikes.

Budget for the peak, too. Per Retell's pricing page, the first 20 concurrent calls are included, additional capacity is $8 per concurrent call per month and can be added or removed as volume changes, and Enterprise plans offer custom concurrency. The Get Concurrency API reports how much more concurrency your workspace can still purchase, so check that ceiling before surge season, not during it.

If you bring your own carrier over SIP trunking, its channel limits and call setup rate count as well. The tightest limit anywhere on the call path is your real capacity.

What happens when calls exceed your voice AI capacity?

On Retell, an inbound call over the limit waits about 40 seconds for a free slot. If none opens, Retell transfers the caller to the phone number's fallback number, or ends the call with concurrency_limit_reached if no fallback is set. An outbound call over the limit is rejected unless concurrency burst is turned on.

Concurrency burst is the short-term buffer. When it is on, calls above your standard limit still go through, up to a burst ceiling equal to the lower of 3 times your limit or your limit plus 300. A limit of 50 bursts to 150. A limit of 200 bursts to 500. Burst calls carry a $0.10 per minute surcharge for the entire call. Retell's docs also note that steady use above your normal limit is a sign to raise the base limit instead.

Your surge plan should record three decisions:

• Burst on or off. On means fewer lost calls during a spike, at a known surcharge. Only the Admin role can turn it on, so make sure the person on call has that access.
• A fallback number on every inbound number. Without one, an overflow caller simply hears the call end. Point it at a human queue, an overflow partner, or a voicemail line that someone actually works.
• Reserved inbound concurrency if the workspace also runs outbound. It caps outbound traffic at your limit minus the reservation. With a limit of 100 and 20 reserved, outbound can use 80 slots and inbound can still use all 100. Batch calls have their own setting that holds slots back for other calls.

One more limit is easy to miss: billing. On credit-based billing, calls stop when the credit balance reaches zero (credit_exhausted). A workspace that reaches its monthly budget pauses calls (budget_reached). A surge spends credits faster than a normal week, so turn on auto recharge and review the budget before peak season.

Which signals tell ops when to add capacity?

Watch calls active against your limit, minute by minute, along with the disconnection reasons that show callers being turned away. A weekly call total cannot tell you when to scale.

Retell alert rules send an email or a webhook (which can post to an on-call channel), can check a window as short as 1 minute, and notify once when an incident opens and once when it resolves. A workspace can have up to 10 rules, so spend them on the signals above. Set the concurrency alert below your limit, around 80% for example, not at it. By the time you hit the limit, callers are already waiting.

How do you check that response time and quality hold up at peak?

Measure them during your own load test instead of accepting a vendor's number. Response time under load depends on your prompts, tools, and backend systems as much as on the platform.

On Retell, call history and the Get Call API show end-to-end latency for each call, including median and P90, and the analytics dashboard charts latency over time. Watch the Custom function latency alert too. During a surge, your own CRM or order lookups often slow down first, and the caller hears that as a slow agent. Compare those numbers from a quiet hour with the peak of your test. If they drift, fix the slow dependency before surge season.

How should overflow and after-hours calls be routed?

Decide which system answers first, then where each kind of overflow goes. Three patterns cover most contact centers:

• AI first, people for exceptions. The voice agent answers every call and transfers the cases it should not handle. In a surge, the AI takes the repetitive calls (status, outage updates, billing dates) and the human queue only receives transfers. Retell's call transfer supports cold, warm, and agentic warm transfers, and a warm transfer lets the agent brief the person who picks up.
• People first, AI as overflow. Your contact center platform keeps call control and sends calls to the AI when queues pass a threshold or a call rings too long. That rule lives in your platform, so the platform's capacity and transfer behavior have to be part of your test.
• AI after hours. Route by time of day. On Retell, the inbound webhook can pick a different agent for each call or reject calls you cannot serve, and conversation flow agents can branch on time variables such as {{current_time}}. If your contact center platform holds call control, its schedule can hand calls to the AI after hours instead. Retell's Five9 guide describes exactly that: live agents during business hours, a Retell agent after hours.

Whichever pattern you choose, map the last hop. Know where a call goes when the AI is full, when the human queue is full, and when a transfer fails. A dead end at any of those points is where surge callers get lost.

Plan what the agent says, too. Surge calls are often about something new, such as an outage or a policy change. On Retell you can assign phone numbers to an environment tag and point that tag at a different agent version, so a pre-tested surge version goes live without editing each number.

What should you test before the surge?

Test what happens to the call over the limit, not just the calls under it. Most load tests prove that calls connect. Fewer prove what the next caller hears once capacity is full, or that finished calls release their slots.

Pre-surge checklist:

• Ramp test traffic through your real call path (carrier, contact center platform, transfer into the AI) past your expected peak. See voice AI load testing that follows the real call path.
• Push past the concurrency limit on purpose. Confirm the fallback number answers, and that burst behaves as planned if it is on.
• Run a closure test so every way a call can end frees its slot. See the call closure test.
• Check calls per second. A burst of calls at the top of the hour can hit a CPS limit before it hits a concurrency limit.
• Trigger every alert rule and confirm it reaches the on-call person. The Test button next to each webhook URL sends a sample payload.
• Pause outbound campaigns, or reserve inbound concurrency, for the surge window.
• Confirm auto recharge and the monthly budget leave room for peak spend.
• Run your surge scenarios (outage questions, billing questions, transfers) as a batch of simulation tests after every prompt change.
• Confirm who on call can change limits: Admin or Developer to adjust concurrency, Admin to turn on burst or reserve inbound capacity.

What should you ask a voice AI vendor about surges?

Ask questions that need a specific answer, not a promise:

• What is my concurrency limit, and does it apply per account, per workspace, or per number?
• What happens to the first inbound and outbound call over the limit? How long does it wait, and where does it go?
• Is there a temporary burst option? What does it cost, and who can turn it on?
• How fast can I raise the limit, and what is the ceiling before I need a contract change?
• Which metrics can I alert on, how short can the window be, and can alerts reach a webhook?
• Can I reserve capacity for inbound calls while outbound campaigns run?
• What stops calls for billing reasons, and how do I prevent that during a surge?
• What happens to my calls if your platform itself has an incident?

FAQ

What is concurrency in voice AI?

Concurrency is the number of calls a voice AI system handles at the same moment. If 15 callers are talking to your agents at once, that is 15 concurrent calls. It is the unit that sets how much peak volume you can absorb before calls queue or overflow.

How many concurrent calls can AI voice agents handle?

As many as the platform's concurrency limit allows. On Retell, Pay-As-You-Go workspaces start at 20 concurrent calls, more can be added at $8 per concurrent call per month, and Enterprise plans offer custom concurrency. Burst can temporarily raise the ceiling to the lower of 3 times the limit or the limit plus 300.

Can AI voice agents answer multiple calls at once?

Yes. A single Retell agent can take an unlimited number of calls at the same time, as long as total active calls stay within the workspace's concurrency limit.

Can voice agents work across multiple support queues at once?

Yes. All agents in a Retell workspace share that workspace's concurrency limit. Separate workspaces have separate quotas, burst settings, and reserved inbound capacity, and traffic in one does not use slots in another. Put queues that must never compete for capacity in separate workspaces.

What happens when voice AI reaches its concurrency limit?

On Retell, inbound callers wait about 40 seconds for a slot, then go to the phone number's fallback number, or the call ends with concurrency_limit_reached if none is set. Outbound calls are rejected unless concurrency burst is on.

How does burst capacity work in voice AI?

Burst lets calls above your standard limit go through during a spike. On Retell, the burst ceiling is the lower of 3 times your limit or your limit plus 300, burst calls cost an extra $0.10 per minute for the whole call, and only an Admin can turn it on.

How do I know when to add voice AI capacity?

Alert on concurrency used at around 80% of your limit and on call volume rising sharply against the previous hour, and track calls ending with no_concurrency_fallback or concurrency_limit_reached. If you rely on burst week after week, raise your base limit.

How does Retell AI handle peak call demand?

Through a per-workspace concurrency limit you can raise from the dashboard, optional burst capacity for short spikes, reserved inbound capacity so outbound campaigns cannot crowd out callers, a fallback number for calls that still overflow, and alert rules on concurrency, volume, transfers, and quality. Retell's reliability overview covers the platform side.

Build the surge plan before you need it

A surge plan is a one-page document: your peak concurrency and the math behind it, your burst and fallback decisions, the alerts and who receives them, your overflow and after-hours routing, and the date you last tested all of it. If you are planning voice AI for a contact center that sees outages, billing cycles, or seasonal peaks, see how Retell supports call center automation, or talk to our team about sizing concurrency for your busiest hour.