Reliability

Reliability

What happens to a call when a check or a provider fails, what is measured, and the targets the platform works to.

The platform's priority order is reliability first, then call quality, then latency.

Answering a call

Several checks run before a call is answered. A billing or warm-up failure never stops a call being answered.

Call arrives
  1. number routing   lookup fails, no recent match -> refuse the call
  2. billing check    error                         -> answer the call anyway
  3. capacity check   no free line                  -> answer, then transfer or apologise
                      concurrent-call limit reached -> decline the call
  4. warm-up          not finished                  -> answer the call anyway

Number routing

The dialled number is resolved to an agent. If that lookup fails, the platform uses the agent it last routed that number to, when that was within the last day. With no such match, the call is refused before it is answered rather than being routed to a default agent.

Billing check

Fails open. A billing error never prevents a call being answered. A billing decision can still decline a call before it is answered: when the workspace's subscription is cancelled, or when the workspace turned on blocking at its spend limit and reached it.

A failed payment never declines a call. The line keeps answering while a payment is overdue, and for 14 days after a subscription ends because a payment failed. The team is alerted when a payment fails, and when the billing check turns away a call to a workspace that has been taking real calls.

Capacity check

If no line is free to take the call, the call is still answered:

  • With an overflow forward number set: the caller hears "One moment while we connect you" and the call is transferred to that number.
  • Without one: the caller hears a short apology and the call ends.

If the platform is at its limit on concurrent calls, or your workspace is at its limit with burst turned off, the call is declined before it is answered.

Warm-up

As soon as a call is placed on a server, the platform starts loading the agent settings and connecting speech recognition and the voice. Loading the settings has a two-second limit. The call is answered without waiting more than a second for this work.

The AI model's prompt cache warm-up and the pre-rendered greeting run in the background and never hold the caller. If the pre-rendered greeting is not ready in time, the greeting is spoken live.

Failover

A failed AI model turn is retried, or switched to a second provider where the workspace allows overseas fallback. Voice failover is available for agents with a fallback voice set.

Failover covers each subsystem in detail.

Self-healing

Each call runs in its own process, which ends with the call. A server that stops answering its health checks is replaced automatically.

Deployment

A build reaches production only after it has passed its smoke tests in our development environment and a set of live test calls. A release whose smoke tests or watched alarms fail in production is rolled back automatically.

Latency

The target is under 500 milliseconds from a caller finishing speaking to hearing any audio in response.

Per-turn latency is recorded for every call and displayed in the product. Latency depends on the model and voice you chose, so measure your own agent.

Backups and recovery

The database is managed by our hosting provider, with point-in-time recovery for the last seven days. Backups are kept in Sydney. Secret values are encrypted with a separate key before they are stored, and are covered by the same backups.

On this page