Voice agent downtime: the fallback and continuity checklist

Four suppliers stand behind one voice

A voice agent handling a single call usually depends on four independent services: the telecom operator that owns the number and carries the audio, the speech recognition that turns audio into text, the language model that decides what to say, and the speech synthesis that turns text back into audio.

Each has its own status page, its own maintenance windows and its own definition of what counts as an incident. None of them can see the other three. The caller hears all four as one thing.

So the question “is the service up?” does not have one answer. It has four, and all four can be yes while the call is unusable.

The most common failure does not look like a failure

When I opened the ElevenLabs status page while writing this article, one incident was open on it: elevated latency on speech-to-text and text-to-speech. It was in monitoring and had been running for six hours. Not down. Slow.

That is the failure most setups handle worst, precisely because nothing is triggered. The agent answers. It just takes a second and a half too long on every turn, and in a phone call a second and a half is a long time. The customer starts repeating herself, the agent hears the repetition as a new statement, and the call stalls in a way that looks like a bad agent rather than a slow supplier. Settling which of the two it was, after the fact, depends on being able to retrieve the record of the call.

It hits the invoice, since the price of a voice agent is usually measured in minutes and the slow calls are the long ones. And it hits the numbers: an agent that is slow but up still scores well on answer rate and on share of calls handled without transfer — the two metrics that flatter most, and the only ones moving in the right direction while the customer experience falls.

The checklist

Six questions. Ask them internally first; the last three take under an hour to answer.

Question	A usable answer
Where does the call go if the agent does not answer at all?	A number or a queue, configured on the phone number itself.
What counts as “down” in the contract?	A latency threshold, not just up or down. Without one, being slow is free for the supplier.
Who at your end notices, and how fast?	A name and a number of minutes. “We watch the status page” is not an answer, and nobody does it at two in the morning.
Who can switch the agent off?	At least two people, and they have to manage it without calling the supplier first.
How long does switching it off take?	A measured figure. Try it on a Tuesday morning and time it.
What does the agent say when it is half down?	A sentence you wrote in advance that hands over to a human. The supplier's default is rarely in your language and never in your tone.

The last row is the one most often left empty. It belongs with the escalation path: an agent that cannot get any further because it is slow should leave the call exactly the way an agent that cannot get any further because the question is too hard leaves it.

The layer below you is regulated. The layer above it is not.

Here is a contrast worth bringing to the continuity meeting.

Take one example from Norway. Nkom, the national communications authority, requires selected base stations in its reinforced-network programme to hold at least 72 hours of backup power during a power cut, against 2–4 hours for ordinary base stations. They must also have a backup connection independent of the main one. The scheme covers 112 municipalities today, with 20 more to be built out in 2026.

Somebody, in other words, has thought carefully about the robustness of the network that carries your call. Nobody has thought equivalently about the layer above it: the model, the speech synthesis and the supplier that ties them together. Up there, robustness is not a regulatory requirement but something you negotiate — or fail to negotiate. It is also the supplier who owns the clock when something gets replaced.

The fallback belongs on the number

That is the one technical detail worth taking away.

Twilio, for instance, documents a separate fallback URL on the phone number itself, which is called when the ordinary handler fails and which carries an error code describing what went wrong. The point is not the supplier. The point is where the fallback sits: a fallback programmed inside the agent does not exist on the day the agent is the problem.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. The best continuity plan for a phone number is a boring one: somewhere for the call to fall back to, two people who can switch it off, and a measured figure for how long that takes. All three can be written down today, and none of them requires anything to go wrong first.