Outage fallbacks · Research · 18 Nov 2026

A fallback for when ElevenLabs goes down

What failed on the Handyman Bury line, how other companies keep a voice agent line answering through an outage, which fallbacks are worth having, and what each one costs.

Recommendation

The short answer

Put our own relay in front of each agent with an 8-second ring timer. If the voice agent has not answered by then, the caller goes to a backup. A standby agent on a second platform comes later, and Retell is the one to trial.

  • The findingElevenLabs did not refuse the Bury call. It left it ringing for 19 seconds until the caller hung up. Telnyx's free failover only reacts to a refusal, so it would not have helped.
  • Do firstAdd the ring timer to the relay we already have (execution/call_desk/app.py). It costs about $0.002 a minute and works call by call.
  • Do nextA watchdog that places a test call on a schedule and messages Telegram when the agent does not answer. ElevenLabs' status page runs up to 2 hours 49 minutes behind.
  • ThenTrial one standby agent on Retell on a demo line. It costs nothing while idle and about $5 for an hour of outage calls.
  • Do not buyElevenLabs Enterprise for its SLA, a human answering service, or Vapi.

From our own Telnyx call records

What happened on 18 Nov

ElevenLabs opened an incident at 15:34. Our call to the Bury line at 15:58 was never answered. By 17:54 the line was connecting again, with the incident still open.

18 November 2026, times in UTC (the same as UK time)

09:00 11:00 13:00 15:00 17:00 19:00 Five calls connected 09:06 to 13:20, another demo line 15:34 ElevenLabs opens its incident 15:58 Bury call never answered rang 19 seconds, then the caller hung up 17:33 ElevenLabs admits inbound SIP calls are failing 17:54 and 18:21 Bury connects again still open 18:33
Green dots are calls that connected. The shaded band is ElevenLabs' incident. Two Bury calls connected inside it, so the fault came and went: it was not a clean on and off.

Why Telnyx's free failover would have missed it

An outage where ElevenLabs refuses the call

Caller rings the line
Telnyx sends it to ElevenLabs
ElevenLabs refuses it
Telnyx tries a second route, if one is set

18 November: ElevenLabs lets it ring

Caller rings the line
Telnyx sends it to ElevenLabs
It rings. Telnyx reads that as a normal call
Caller gives up after 19 seconds
Telnyx only moves a call on when it is refused or the network fails. Ringing with no answer "is not considered a failure" (Telnyx failover; Telnyx call forwarding).
  • The Bury line has one route and nothing behind it. One address (sip.rtc.elevenlabs.io), no second route, no call forwarding. A failed call has nowhere to go.
  • ElevenLabs has no setting for this. Its only resilience option is a backup language model, which does nothing when the call never starts (ElevenLabs docs).
  • One caveat. This rests on a single call record. Other outages may refuse calls outright, and a second route would catch those.

How often, and what they owe us

ElevenLabs' record

This will happen again. The fallback will be used regularly, not once.

  • About 13incidents in 2026 where phone calls failed to connect or dropped, in ten and a half months. Five were in the EU only.
  • 11h 37mof telephony trouble posted in the last 86 days, about 99.4% clear. 18 November is not in that figure: ElevenLabs did not tag it as telephony.
  • About 12 hourswas the worst outage, on 31 July, in the EU and India. The US was down longer.
  • 35m to 2h 49mis how long after the real start ElevenLabs first posts. The status page is too slow to be our alarm.
  • Noneis the uptime promised on our plan. The service is "as is". An SLA exists only on Enterprise, and its terms are not published.

"Resolved" has also relapsed inside the same incident at least three times this year, so a fallback should stay armed after the all-clear. (ElevenLabs incident data; third-party archive; terms)

The pattern others publish

What other companies do

They build a ladder. Each rung catches a different failure, and the last rung needs no AI provider at all. Twilio published the clearest version on 4 November 2026 (Twilio).

RungWhat it catchesWho offers it
Backup speech or language modelOne supplier failing while the platform is upElevenLabs, Retell, Vapi, LiveKit
Fallback number on the platformSending calls elsewhere by handRetell "backup mode", Vapi fallback destination
Carrier rerouteThe platform refusing callsTelnyx second route and forward on failure; Twilio equivalents
Relay with a ring timerThe platform stalling, as on 18 NovemberTelnyx TeXML and Twilio TwiML
Scheduled test callsKnowing within minutesMonitoring vendors advise every 5 to 15 minutes
  • A fallback run by the platform that is down is weak. The only user report found was a Vapi customer whose fallback number did not fire in a test (Vapi support thread).
  • Published uptime promises cluster at 99.9%. That still allows about 8.8 hours down a year.
  • No agency has written up what it does. Searches of Reddit, Discord and similar found nothing first-hand. A tested fallback with a written procedure puts us ahead of what is public.

From a free setting to a standby agent

Our options

The test for each one: would it have caught the 18 November call?

OptionCaught 18 Nov?Standing costCost in useEffort
A. Second route on the Telnyx connectionNo$0$0Minutes, but needs somewhere to point
B. Telnyx "forward on failure" to a phoneNo$0.20 a number a monthTwo billed call legsMinutes
C. Our relay with an 8-second ring timer, then voicemail and an alertYes$0About $0.002 a minute on every callDays. Extends the relay we have
D. Relay, then the client's staff line, then voicemailYes$0About $0.024 a minute to a UK mobile, $0.009 to a US oneC, plus a number from each client and Roman's agreement
E. Relay, then a standby agent on a second platformYes$0About $0.06 to $0.08 a minute while it answersWeeks, then upkeep
F. Human answering serviceYes, through the relay$49 a month or more$2.99 a minuteA script per client
W. Watchdog: scheduled test call and Telegram alertDetects, does not rescue$0About 9 cents a test call at worstDays. Starts from a script we have

The call path we would build

  1. 1The voice agent on ElevenLabs answers, as now. The relay gives it 8 seconds.Every call
  2. 2A standby agent on a second platform takes the call if the first does not answer.Later, option E
  3. 3The client's own staff line or mobile rings next.Option D
  4. 4Voicemail, with a Telegram alert saying "taken by backup". No AI provider is involved.Option C
Options C, D and E are the same relay with more rungs added. This is the design already written in LEARNINGS on 13 October. One correction: that note says about 4 seconds, but Telnyx's ring timer has a 5-second floor.

Things to know before building

  • A and B still help. They catch outages where ElevenLabs refuses calls. They are just not enough alone.
  • A cost trap on D. If the forwarded leg shows the caller's own number, a UK mobile leg can bill up to $0.835 a minute, not $0.02. Show the business's number.
  • The relay puts our VPS in the call path. Telnyx has a fallback address for when the VPS is down. Agent transfers to a human are the likeliest thing to break behind a relay and need testing on a copy.
  • The silent case stays open. If the agent answers and says nothing, no phone network setting notices. Only the watchdog catches that.

Reliable companies for a second agent

Which standby platform

No platform has had a clean year. A standby earns its place by failing at different times from ElevenLabs, not by a better record.

Retell AI

Trial this one
  • A separate company from ElevenLabs and from Telnyx's AI layer.
  • Documented connection from a Telnyx line.
  • Calmest recent status page: its worst recent event lasted 13 minutes.
  • Against: slowest in one open speed test, and some users report Telnyx setup problems.

$0 idle. About $0.07 to $0.08 a minute. 20 calls at once.

Telnyx AI Assistants

Second choice
  • No second supplier, no extra hop, and the cheapest.
  • Against: it failed on us on 28 July. 17 calls in a row were answered and silent, which a ring timer cannot detect.
  • Against: it shares the carrier with the main path, and voice choices are locked.

$0 idle. About $0.06 a minute. 500 calls at once.

OpenAI Realtime or LiveKit Cloud

Only if we want to own code
  • The most independent of ElevenLabs.
  • Against: we would write and run the agent ourselves.
  • Against: users report failed inbound calls on OpenAI's phone connection.

$0 to $50 a month idle. $0.05 to $0.38 a minute.

Vapi

Ruled out
  • Two long outages on its bring-your-own-line service in October: 12 hours 18 minutes and 7 hours 47 minutes.
  • 41 of 246 test calls never connected in one benchmark.
  • It reported failures across three speech suppliers on 18 November, ElevenLabs included.

$0 to $29 a month idle. About $0.08 to $0.13 a minute.

Also ruled out: Hume EVI (reported retired), Deepgram and Gemini Live (no direct phone connection), Bland and Synthflow (no reliability evidence), and a second ElevenLabs region (Enterprise only, and not a failover).

Three cautions on any standby

  • The voice. A good British voice without ElevenLabs is unproven on every platform. It needs a listening test on a handset.
  • The clinic. If its calls count as health data, a standby needs its own signed agreement first. Vapi charges $2,000 a month for that and Telnyx lists $2,500. Retell says its agreement is self-serve, in its own marketing. Until that is settled, the clinic's fallback is its staff line and voicemail.
  • The upkeep. A second prompt, tools and test pack must be kept in step with the main agents. That time will cost more than the cash.

List prices read on 18 Nov 2026, at £1 = $1.32

Pricing

A standby is cheap because it sits idle. What matters is whether there is a monthly fee, and how many calls it can take at once when every line falls over together.

PlatformMonthly fee, idlePer minute in useCalls at onceFee for our Telnyx line
ElevenLabs, as nowOur plan$0.08 plus the modelBy planNone
Retell AI$0$0.08, or $0.11 with a premium voice20None
Telnyx AI Assistants$0About $0.06500None
Bland$0$0.14 flat10None
Vapi$29 for 10 lines$0.08 to $0.1310None
LiveKit Cloud$50 for 20 linesAbout $0.05, plus our own code20Included minutes

Added cost of a full-day outage: 500 call minutes across 10 lines

  • Forward to a US mobile$3 to $7 £2 to £5
  • Forward to a UK mobile$10 to $12 £8 to £9
  • Telnyx AI Assistants$30 £23
  • Retell AI$41 £31
  • Bland$70 £53
  • LiveKit Cloud$74 £56
  • Vapi$70 to $93 £53 to £71
  • Human answering service$1,544 £1,174
The striped bar is not to scale: the answering service (PATLive, the cheapest pay-when-used plan found) costs about 37 times as much as Retell for the same day. Vapi and LiveKit include their monthly fee. Our numbers and inbound minutes cost the same whichever agent answers, so they are left out.
Other costsPriceNote
A two-hour outage on Retell, about 60 call minutes$5 (£4)$3.60 on Telnyx AI Assistants
The relay itselfAbout $0.002 a minuteTelnyx's fee per call leg, on every call
Watchdog test calls$65 a month hourly, $260 every 15 minutesWorst case, if each test bills a full ElevenLabs minute at the overage rate
Text to a callerUK $0.05, US about $0.009Telnyx's own pages disagree on the UK rate
Bought-in test call monitoring (Cekura)$0.25 a test minuteNo monthly fee. Others are quote only

Recommendations

What to do, in order

  1. 1

    Add the 8-second ring timer to the relay and put the Bury demo line behind it. The backup is voicemail plus a Telegram alert. Test it against a dead address.

    CostAbout $0.002 a minuteAgreesRachel. It is her connection

  2. 2

    Build the watchdog. A test call to one line on a schedule, plus a free check of ElevenLabs' status feed, alerting on Telegram.

    CostUp to $65 a month hourlyAgreesRachel. Tell Roman, so his dashboards skip the test calls

  3. 3

    Take the tested design to Roman for the live client lines. Agree for each client what the backup is: their staff line, then voicemail.

    Cost$0.002 a minute, plus forwarded minutes in an outageAgreesRoman and Rachel both. The lines are on his connection

  4. 4

    Check the backup language model is switched on for each live agent. Looking is read-only.

    Cost$0AgreesNobody to look. Both to change anything

  5. 5

    Build one standby agent on Retell for a demo line. Run the test pack and a listening test. Extend to live lines only if it passes.

    Cost$0 idle, about $5 an outage hourAgreesRachel to trial. Both for live lines

  6. 6

    Promise clients what we control. "A caller always reaches something": the voice agent, then their own staff, then a voicemail that alerts them. Not an uptime percentage.

    Cost$0AgreesRoman and Rachel

Do not buy: ElevenLabs Enterprise for the SLA (an SLA pays credits, it does not answer calls), a human answering service, or Vapi. Nothing in this plan releases, ports or deletes a number, and Roman's connection is left alone until he agrees to a change.

Read this before acting

What is not verified

  • Nothing here has been tested with a real call. Every claim about fit comes from documentation.
  • Whether ElevenLabs sent a "ringing" signal before stalling, and what it returns in other outages.
  • The cause and full length of the 18 November incident. It was still open at 18:33.
  • Whether the relay can reach ElevenLabs over the connection type ElevenLabs requires. Test this first.
  • The exact cost of the relay's second leg, and UK per-minute rates on our account.
  • Whether Retell's health-data agreement is really free and self-serve.
  • Twelve-month incident totals for any platform. Official pages keep 90 days and trackers disagree.
  • What other agencies do. No first-hand account was found.

Where the figures come from

Sources