Call Centre Infrastructure for AI: What Happens When the AI Is Down

Research date: 23 September 2026. Last updated: 23 September 2026.

Put an AI voice agent in front of your phone line and you have not added a feature. You have added a dependency, and your phone line is now only as available as the AI is.

That matters more than it sounds, because the two layers you are stacking make very different promises. Amazon Connect commits contractually to 99.99% monthly uptime with a published schedule of service credits. Vapi's own status page reports 99.890% on its API and 99.912% on its SIP service, and states no contractual commitment at all. One of those is a promise with money behind it. The other is an honest description of what happened last month.

Converted to minutes, the gap is not subtle. Against a 30-day month, 99.99% is about 4 minutes of downtime. 99.89% is about 48 minutes — and when the thing that is down is the system answering your phone, those are minutes in which nobody is answering it.

Brilo AI publishes this page and sells an AI voice agent. Brilo's own terms state that it "does not ensure that the service will be continuous, on time, secure, or without errors", and publish no uptime commitment. That is quoted again below rather than buried here.

How we compiled this

Documented research. Uptime figures, SLA terms and status pages were read from the vendors' own sites and legal documents on 23 September 2026, and HTTP status checks were run against the conventional SLA paths. Nothing here is a measurement of our own.

The three ways an AI agent reaches your phone line

Most of the infrastructure decision comes down to which of these you pick, and vendors rarely lay out the trade-offs.

Number forwarding. You keep your existing number and conditionally forward it — all calls, after hours, or on busy — to a number the AI provider owns. Fastest to set up, often same-day, and reversible in minutes by cancelling the forward. The costs: you are adding a hop, your carrier's forwarding rules are now part of your call flow, and caller ID handling varies by carrier.

SIP trunking. You connect your telephony to the AI provider over SIP rather than bouncing a call through the PSTN twice. Lower latency, better control of call data, and the route that suits anyone with existing SIP infrastructure. The cost is that somebody has to configure it, and SIP misconfiguration produces the kind of one-way-audio faults that take days to find.

Native CCaaS integration. Your contact centre platform offers the AI as a component inside its own routing. Cleanest operationally, because escalation and reporting stay in one system. The cost is that you are limited to what your platform partners with, and you have made your platform harder to leave.

The decision rule. Forwarding to evaluate, because reversing it is trivial. SIP or native once you are committed and latency and data control start to matter. Do not start with SIP to save a migration later — you will spend the first two weeks debugging audio rather than the conversation.

The availability asymmetry, in figures

This is the part that should shape the contract.

Sources: aws.amazon.com/connect/sla, twilio.com/en-us/legal/service-level-agreement, status.vapi.ai, status.retellai.com, brilo.ai/terms — all read 23 September 2026. HTTP 404 returned on vapi.ai/sla, retellai.com/sla, deepgram.com/sla and elevenlabs.io/sla on the same date.

Two things worth saying precisely.

A status page is not an SLA. It is a record of what happened, published voluntarily, and it can be accurate and admirable — Vapi's is both — while committing the vendor to nothing. An SLA is a contractual promise with a remedy attached. Confusing the two is the most common mistake in this evaluation.

And the contractual one has conditions. Amazon's 99.99% credits require an Enterprise Support plan and a completed Well-Architected Review with an AWS Solutions Architect, with claims filed within two billing cycles. A commitment you cannot practically claim against is worth less than its headline.

The question nobody asks: what happens to the call?

Availability figures are abstract. This is the concrete version, and it is the single most important thing to settle before you go live.

When your AI provider is down, what happens to a caller dialling your number? There are four possible answers and they are wildly different:

• The call falls through to your human queue as though the AI were not there. Best case, and it is a configuration you have to ask for.

• The call rings out unanswered.

• The caller gets a busy tone or a carrier error.

• The call connects and then sits in silence, which is the worst of the four because the caller believes they have reached you.

If you are on number forwarding, you have a real advantage here: conditional forwarding can be reverted, and some carriers support a failover destination if the forward target does not answer. If you are on SIP or native integration, the failover behaviour is whatever the platform does, and you need it in writing.

Test it before you launch. Ask the vendor to simulate an outage, or point the forward at a dead number for five minutes during a quiet period and ring in yourself. What you hear is your actual failure mode. Almost nobody does this, and it is a fifteen-minute test.

Centralising comms, and when not to

The other half of this question is consolidation: several numbers, sites or systems being pulled into one. AI is often the reason it finally happens, because a voice agent needs one coherent place to route from.

Consolidate when you have numbers nobody owns, when callers get different answers depending on which site picks up, or when you cannot report on total call volume because it lives in three systems.

Do not consolidate at the same time as deploying AI. Two changes to the same phone estate in the same fortnight makes every fault ambiguous, and you will spend the first outage arguing about which project caused it. Consolidate, let it settle, then add the agent — or add the agent to one line, prove it, then consolidate around what worked.

Keep a number that bypasses everything. One direct line to a human, known to your team and given to anyone who has already had a bad experience. It costs nothing and it is the difference between a complaint and an escalation.

What to ask before you sign

1. Do you publish an uptime SLA, and what are the service credits? If the answer is a status page, say so plainly.

2. What happens to an inbound call when your service is unavailable — falls through, rings out, busy, or silence?

3. Can I configure a failover destination, and will you help me test it before launch?

4. Which connection methods do you support — forwarding, SIP, native integration — and what is the latency difference?

5. If you rely on third parties for speech, models or telephony, does your availability depend on theirs, and is that reflected anywhere in what you commit to?

The last one matters because most voice products are assembled from components. Vapi's own status page lists OpenAI as a monitored dependency, which is candid — but it also means the availability you experience is a product of several services, not one.

Brilo's position, and its own terms

Brilo publishes a free tier at 10 minutes a month, Pro at $149 for 600 minutes with overage at $0.16, Growth at $499 for 2,500 minutes at $0.14, and a custom tier above 5,000 minutes. Source: brilo.ai/pricing, read 23 September 2026.

What it does not publish: Brilo has no SLA page — brilo.ai/sla returns HTTP 404 — and no uptime commitment or service credits anywhere in its terms. Its terms state that Brilo "does not ensure that the service will be continuous, on time, secure, or without errors", provide the service "as is", and disclaim responsibility for damages arising from interruptions or service failures. Source: brilo.ai/terms, read 23 September 2026.

That language is ordinary for self-serve software and unremarkable in itself. It is worth knowing anyway, because the argument of this page is that an AI agent in front of your phone line inherits its availability to you — and on a published-terms basis, Brilo currently commits to none. If uptime is contractually important to your operation, ask what is available on a custom plan and get it in writing, and apply the same question to every vendor here.

Related reading

For whether the deployment is working once it is connected, see AI call centre quality assurance. For capacity and concurrency rather than availability, see AI workforce management and diallers. For the deployment sequence itself, see how to implement AI in your call centre.