What Happens If Your AI Receptionist Goes Down? A Backup Plan for Every Call (2026)

Short answer: If your AI receptionist goes down, the call either rings out, hits an error message, or ends, unless you set a backup first. A good backup plan covers five different failures: too many calls at once, a platform outage, your own webhook failing, the voice model failing, and the call summary never reaching you. On Retell, that means a fallback number on every line, Backup mode you know how to switch on, a default agent on the number, and alerts on the workflow that delivers your messages. Outages are rare but real. On the client lines we run, 436 of 437 calls from October 15 to November 13, 2026 ended normally, and the plan exists for the one that didn't.

We build AI voice receptionists on Retell, with n8n sending each call summary to the owner. Here's what can break, what each vendor does about it, and the backup we set up before a line goes live.

What happens to a call when an AI receptionist fails?

It depends on which part fails, and each failure has a different fix.

What fails	What the caller gets with no plan	The backup	Where you set it
Too many calls at once	Waits about 40 seconds, then the call ends	Fallback number (staffed line or voicemail)	Retell phone number settings
The platform has an outage	Ringing, silence or a dropped call	Backup mode, or forwarding from your main line	Retell phone number, your carrier
Your inbound webhook is down	Call drops if no default agent is set	Keep a default inbound agent on the number	Retell phone number
The voice model or its provider fails	Silence or a failed call	A backup model, or a voice setup that isn't tied to one provider	Agent settings
The call works but the message never arrives	Caller thinks they're booked; you never hear about it	Webhook retries plus an error alert	Your webhook and automation tool

Too many calls at once: set a fallback number

Every Retell workspace has a concurrency limit, the number of calls it can handle at the same moment. Retell's docs say Pay-As-You-Go workspaces get 20 concurrent calls by default. A one-truck plumber won't hit that. A workspace that also runs outbound reminder calls can.

When inbound calls reach the limit, Retell holds new callers for a slot. If no slot opens after about 40 seconds, Retell transfers the caller to the phone number's fallback_number. Without one, the call ends with concurrency_limit_reached. Two rules from Retell's API reference matter when you pick the fallback: it can't be the same number, and it can't be a number that already has its own fallback (Retell blocks nested forwarding).

Two more settings help:

• Reserved inbound concurrency holds part of your limit for incoming calls, so an outbound campaign can't crowd out customers calling in.
• Concurrency burst lets calls go over the limit for an extra $0.10 per minute, charged on the whole call.

A diverted call has no agent fee; on Retell-managed numbers you pay telephony from when the fallback answers (Retell's example: $0.15 for 10 minutes at $0.015 a minute).

Our build rule: the fallback must be a line a person answers, or a voicemail someone checks, and never the number forwarding into the AI, or the caller loops. The same rule applies to transfers, which we cover in how an AI receptionist transfers to a human.

When the platform itself is down

Retell's fallback number is an overflow setting. Retell's docs describe it for full concurrency and for Backup mode, a switch on each phone number that sends every new inbound call straight to the fallback number before an agent is picked. You turn Backup mode on yourself (it stays disabled until a fallback number is saved), and you turn it off to go back to the AI. If Retell can't connect to the fallback destination, the call ends.

So for a real platform outage, you need two things: a person who knows where the switch is, and a route that doesn't depend on the AI platform at all. That second route is usually your own carrier. If your main business number forwards to the AI only when you don't pick up, your phones still ring first. If it forwards every call unconditionally, the AI platform is the only path in.

Outages do happen. Retell's public status page lists, among others (times converted to Eastern):

• July 29, 2026, 10:55 to 11:42 AM ET: "Inbound calling not connecting" for certain numbers.
• September 5, 2026, 7:10 to 8:19 PM ET: web and outbound calls failing, inbound calls delayed.
• September 16, 2026, 9:15 to 9:35 PM ET: a call service disruption.

That's normal for a phone platform, and it's why the plan has to exist first. Other platforms handle it differently:

• Twilio numbers have a Voice Fallback URL ("Primary handler fails"). If Twilio can't get instructions from your server, it requests the fallback URL right away, which can simply redirect to a support line. Call webhooks have a hard 15-second limit.
• 3CX added this on November 11, 2026 in V20 Update 10: if the AI provider has an outage or the API key fails, calls go to a backup destination you choose, such as a support queue.

When your own webhook fails

Many builds use Retell's inbound call webhook to look up the caller or pick an agent before the call connects. Retell gives that webhook 10 seconds per attempt and makes up to 3 attempts. If all three fail, Retell checks whether the number has an inbound agent set. If it does, the call connects to that agent. If not, Retell disconnects the call.

The cheap fix: always leave a default inbound agent on the number, even when a webhook normally picks one.

When the voice model fails

From 4:45 PM ET on October 24 to 7:50 PM ET on October 25, 2026, Retell's status page reported that a permission issue with OpenAI's speech-to-speech model caused calls on speech-to-speech agents to "fail or go silent." Only those agents were in its path. The lines we run use a separate voice (ElevenLabs and Retell voices) rather than a speech-to-speech model.

This matters right now because Retell launched GPT-Live-1 on November 2, 2026, a full-duplex voice model for single-prompt agents on paid accounts. Retell's SDK changelog lists an optional backup in that setup. If you switch to it, set the backup and test what callers hear when the primary model is unavailable. We compare voice setups in AI receptionist voice models for 2026.

The quiet failure: the call works, the message doesn't

The receptionist takes a perfect message, then the handoff to your phone, CRM or inbox breaks, and nobody knows. This is the one we watch most closely.

Retell's call webhook has a 10-second timeout and retries up to 3 times when your endpoint doesn't return a success code. Retell warns that you may get the same event more than once, so the workflow that receives it should ignore duplicates. Even so, if the endpoint stays broken, the message is lost.

Here's what we run on our lines:

• Error alerts on the delivery workflow. On November 2, 2026, we attached an error workflow to the n8n workflow that logs every call on our client lines and to our own Retell backend. If a run fails, an alert email goes out instead of the failure sitting in a log.
• A daily call review. Every day we pull the calls on each line, separate real callers from spam, and flag repeat callers within 14 days. A customer calling back twice is often the first sign that a message didn't reach the owner.
• A spot check against Retell's call list. From October 15 to November 13, 2026, the client lines we manage took 437 inbound calls. 436 ended normally (the caller hung up, the receptionist ended the call, the call was transferred, or it timed out on silence). One ended with a platform error on October 28, less than a second into the call.

We cover tool choices in Retell workflows vs n8n and Zapier and what a complete message looks like in AI receptionist message taking.

AI receptionist backup plan checklist

• A fallback number on every line, answered by a person or a voicemail someone checks
• Fallback is not the number that forwards into the AI
• Someone knows where Backup mode is and has tested turning it on and off
• Your main number forwards on no answer, not every call, if you want your phones to ring first
• A default inbound agent stays on the number, even when a webhook picks the agent
• Reserved inbound concurrency set if the workspace also makes outbound calls
• A backup model set and tested if you use a speech-to-speech or duplex voice
• Error alerts on the workflow that delivers call summaries
• A daily or weekly call review that flags repeat callers
• The same backup plan checked before closed days (holiday hours guide)

Want us to check your setup?

We'll look at every line you run, set the fallback numbers, test Backup mode and the webhook failure path, and put alerts on the workflow that delivers your messages. See Retell AI voice agent setup, Retell AI integrations and webhooks and missed call capture, or book your AI systems assessment.