Your AI receptionist’s uptime number is printed on a pricing page. That is not where a promise lives.
If the platform behind your AI receptionist goes down, your phone stops ringing and nobody owes you the job you lost. I read the pricing pages and the public terms of three AI voice platforms on 23 September 2026. Only one of the three puts an uptime figure inside the contract you actually sign, and none of the three publishes a service credit for missing it. The layer underneath them, the telephony provider, publishes both, and that is where the protection you can actually buy is sitting. I build AI calling for home services companies, so I read these documents with a plumber’s Saturday in mind.
The uptime number on a pricing page is a marketing asset. The one inside the terms is a promise.
The mistake: reading uptime off the pricing page
Buyers ask me three questions about an AI receptionist: cost per minute, how many calls it takes at once, and whether the voice passes on a demo. I am almost never asked what happens when the platform is unavailable.
The reason is that uptime looks settled. A figure sits on the pricing page, carries two nines or three, and reads like a commitment. It is a sentence on a marketing page. Whether it binds anybody depends on a different document, and on two of these three platforms that document says something else.
What three AI voice platforms actually promise
Everything below was read from the vendors’ own pages on 23 September 2026.
Bland publishes the best number and the thinnest document. Its pricing page carries an Uptime SLA row in the plan comparison reading 99.9 per cent on Start, on Build and on Enterprise. Its public terms of service page contains no occurrence of uptime, no service credit and no 99 anywhere in it. What it does contain is a disclaimer that the service is provided AS IS without warranties of any kind, and a limitation of liability stating that your sole remedy for dissatisfaction is to stop using the service.
Uptime SLA 99.9 per cent, on all three plans.
Read from the plan comparison table on 23 September 2026.
No uptime clause, no service credit, the service provided “AS IS”.
Read from the public terms of service page the same day.
Vapi puts uptime behind a paywall. Its pay as you go tier lists email support without an SLA. The $29 Core package buys a two business day response SLA and still no uptime commitment. The first uptime figure anywhere on the ladder sits on Pro, which is priced at 10 per cent of the Vapi hosting fee with a $999 a month minimum, and that figure is 99 per cent. The 99.9 per cent only arrives on Premier, which is contact sales. Vapi’s terms of service name no uptime figure either. They disclaim any warranty that the service will be uninterrupted, except as provided under a service level agreement, and I could not find that agreement published anywhere on the site.
Retell inverts the pattern, which is why I stopped trusting pricing pages for this question. Its pricing page names no uptime figure at all. Section 12 of its terms of service does: Retell will use commercially reasonable efforts to maintain at least 99.5 per cent uptime, excluding scheduled maintenance windows not to exceed six hours per month, and force majeure. That is the only one of these three numbers I found inside a contract, and it belongs to the vendor whose pricing page advertises no uptime at all.
What those percentages buy you in minutes
A 30 day month holds 43,200 minutes, and that is the denominator behind every one of these percentages.
Run them. Vapi’s 99 per cent, the cheapest uptime commitment on that ladder at $999 a month, permits 432 minutes of unavailability, or 7 hours 12 minutes. Retell’s 99.5 per cent permits 216 minutes, and the maintenance carve out adds up to 360 more, so about 9 hours 36 minutes can pass without a breach. Bland’s advertised 99.9 per cent permits about 43 minutes. Twilio’s published 99.95 per cent threshold permits about 21 minutes 36 seconds.
Set the two ends against each other. The cheapest uptime commitment you can buy on Vapi allows twenty times more downtime than the telephony layer underneath it, and costs $999 a month to get. And commercially reasonable efforts is an effort standard, not a guarantee, so Retell’s number is a commitment to try.
None of this says these platforms are unreliable. I have no outage data on any of them. It says the documents do not promise what the pricing pages imply, and a contract is the only place a promise survives a bad night.
The layer underneath publishes a real remedy
Twilio’s API service level agreement is the shape these three have not adopted. It names a monthly availability threshold of 99.95 per cent for its services APIs, rising to 99.99 per cent for customers on Enterprise Edition. More to the point, it defines what happens when the threshold is missed: an API service credit of 10 per cent of the fees the customer incurred for the affected APIs in that calendar month.
Two qualifications. The credit covers Twilio’s fees, not your AI vendor’s, and not the job you lost. And the customer is eligible to request it, which means somebody on your side has to notice the outage and ask. Ten per cent of one month of telephony is small money. It is still more than nothing, and nothing is what those three publish.
The second reason the telephony layer matters more is practical. Twilio documents a fallback URL on every phone number, used when the primary webhook fails, and its failover guidance says that fallback can be as simple as redirecting the caller to a support line. Your AI platform is the primary webhook. Something has to sit behind it.
Three moves before your next storm day
Keep the phone number in your own telephony account
The quickest setup is to let the platform hand you a number. If the number lives in the vendor’s account then the failover switch belongs to the vendor. Vapi publishes a guide to importing an existing Twilio number, so this costs you an afternoon, and it leaves you able to point the line somewhere else on a bad morning.
Set the fallback URL on that number
Twilio’s failover guidance is blunt about why: even a highly available application will not survive a data centre failure or a network problem. Anywhere you can specify a webhook URL you can also specify a fallback, and Twilio says that fallback can be as simple as redirecting to a support line. Point it at a mobile that rings.
Ask for the SLA document in writing before you sign
Two questions settle it. Where is the uptime figure defined, and what is the remedy when it is missed. If the answer is a link back to the pricing page, you have just learned what that number is worth. A vendor that answers both in writing has thought about your worst night.
One caveat on the first move. A bundled number is genuinely simpler to run, and simpler is worth something. The trade is that on the worst morning of the year the person who can redirect your calls works somewhere else. I wrote separately about the handoff that breaks when nobody answers the transfer, which is the same failure one layer up.
The bottom line
Uptime is the specification buyers read off a marketing page and rarely verify. Across the three platforms I read this week, the vendor advertising the best number has no uptime clause in its contract, the vendor advertising none has the only contractual one, and not one of them publishes a service credit. I have built this for more than 200 businesses: the protection you can actually buy is a phone number you control and a fallback that takes ninety seconds to set.