Enterprise Voice AI Uptime SLA: What Is Reasonable

Quick answer: A reasonable enterprise SLA for a voice AI platform is 99.9% a month: 43.2 minutes of downtime in a 30-day month. 13 of the 21 plan-level SLA figures we read (28 October 2026, with Telnyx and Vonage added 6 November 2026) sit there (AWS, Google Cloud, Microsoft, OpenAI, Vapi, Bland, Deepgram). 99.99% exists at premium and paid-support tiers, and in Telnyx’s machine-readable SLA file. A 99.95% carrier plus four 99.9% layers gives about 99.55% end to end.

What’s a reasonable enterprise SLA for voice AI uptime?

The honest answer is a range, and the range comes from what vendors put in writing rather than what anyone thinks they should. We read the SLA, legal, pricing and documentation pages of 11 vendors on 28 October 2026 (a twelfth, Vonage, blocked us) and wrote down every availability percentage they publish, which plan it applies to, how they measure downtime and what they pay when they miss. On 6 November 2026 we added two figures: Telnyx’s machine-readable SLA file, which our first reading missed, and Vonage’s support-plans page, read through a text reader. Counting one figure per published SLA line (one line can cover several products or plans, as AWS’s language-services SLA and Bland’s plan table do), and taking the contract wording where a vendor states more than one, gives 21 figures in five bands (Gemini’s 95% figure for models Google designates for shorter availability is shown beside its 99.5% line, not counted separately).

Published monthly uptime	Downtime allowed, 30-day month	Who publishes it (plan or service)	What that tells you
99.99%	4.32 minutes	Amazon Connect Customer (per AWS region); Twilio APIs for Twilio Enterprise Edition customers; Vonage Premium and Enterprise support plans; Telnyx core voice and messaging (machine-readable SLA file, not in its terms)	Available at the contact-centre, premium-edition or paid-support tier; Telnyx states it with no plan named, but in a file its legal terms do not mention
99.95%	21.6 minutes	Twilio APIs (standard)	The communications-API (CPaaS) default
99.9%	43.2 minutes	Google Cloud Speech-to-Text, Text-to-Speech, Dialogflow and Contact Center as a Service; AWS ML Language services (Lex, Polly, Transcribe); Amazon Bedrock; Azure AI Services (which include Speech); Azure Foundry Models (which include Azure OpenAI); OpenAI Fast mode and Scale Tier (Enterprise customers); Vapi Premier; Bland (all three plans); Deepgram Dedicated	The cluster. This is the market’s enterprise number for speech, model and agent-platform layers
99.5%	216 minutes	Gemini Online Inference API (95% for models Google designates for shorter availability); Retell AI terms of service	A contractual floor some vendors write while marketing a higher figure
99%	432 minutes	Vapi Pro	A scaling-team tier, not an enterprise tier

So the reasonable ask, stated as a sourced range: 99.9% a month for the voice AI platform, with a written credit schedule, is the market norm; 99.95% to 99.99% is achievable at the carrier and contact-centre layers; 99.5% is the lowest figure we found in a vendor’s contract terms for a standard service. Below it you are buying a scaling-team tier (Vapi Pro, 99%) or a model Google designates for shorter availability (95%). Retell AI is the clearest example of why the band matters: its terms of service commit to 99.5%, its documentation says “we guarantee >99.9% uptime”, and its enterprise platform page says “Retell offers 99.99 percent uptime”. Each figure is real and each sits on Retell’s own site. The one that governs your contract is the one in the document you sign.

The 99.9% figure is a floor, not a ceiling. Amazon Connect’s 99.99% shows that four nines is commercially available when the vendor runs the contact-centre platform itself, though Connect excludes carrier problems beyond its demarcation point and pays credits only to customers on an AWS Enterprise Support plan who have completed a Well-Architected Review. If your agent runs on a contact-centre platform rather than a developer API, ask for that tier.

Where Zian fits in a voice AI uptime decision

Zian AI is the agent platform in this stack: an AI receptionist that answers inbound calls 24/7 (qualifying, booking, transferring and taking messages) and outbound agents that follow up by voice, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, in 30+ languages. As at 28 October 2026, zian.ai does not publish an uptime SLA, so this page does not claim one; ask for the figure, the measurement method and the credit schedule in your partnership application, and hold Zian to the same voice AI SLA checklist below as every other vendor.

Two published Zian capabilities matter for availability planning. Zian supports private model deployment on customer infrastructure, which moves the model layer inside an environment you control, and it handles answering machines and voicemail on calls. Zian is in a partnership-application beta with no published prices; each of its three listed plans, including Intel + Widgets (Free Forever), starts with a partnership application. If you would rather not run the platform yourself, LeadsNow.ai is the done-for-you, pay-per-result option that runs on Zian, with leads answered in under 60 seconds, 24/7.

What does each uptime percentage allow in downtime?

An availability percentage converts directly into minutes. The table below is the threshold table to hold beside any SLA you are offered. Every figure is computed from 43,200 minutes in a 30-day month, 10,080 minutes in a week and 525,600 minutes in a 365-day year.

Uptime SLA	Downtime per week	Downtime per 30-day month	Downtime per year	Use it as
99.0%	100.8 minutes	432 minutes (7.2 hours)	87.6 hours	Too low for a phone line that books revenue
99.5%	50.4 minutes	216 minutes (3.6 hours)	43.8 hours	Minimum acceptable contractual floor
99.9%	10.08 minutes	43.2 minutes	8.76 hours	Reasonable enterprise ask for a voice AI platform
99.95%	5.04 minutes	21.6 minutes	4.38 hours	Reasonable ask for the carrier or CPaaS layer
99.99%	1.01 minutes	4.32 minutes	52.56 minutes	Premium contact-centre or support tier; where a vendor states it without one (Telnyx’s SLA file), get it into your signed order
99.999%	6.05 seconds	25.9 seconds	5.26 minutes	None of the voice, speech or model services in our tables publishes it

Two caveats change what those minutes mean. First, the measurement interval: Twilio does not count unavailability shorter than five continuous minutes, Google counts Speech-to-Text downtime from one consecutive minute of more than 5% errors, and AWS measures in five-minute intervals. Second, an hour of downtime at 3 a.m. on a Sunday and an hour at 10 a.m. on a Monday cost the same against the SLA and very different amounts in missed bookings. The carrier side of this arithmetic, including a worked Twilio credit, is in our reading of what carriers actually sign in their voice SLAs; this page stays on the platform, speech and model layers above the carrier.

Voice AI platform uptime SLAs, ranked

How we built this table. We wrote these definitions before filling any row and applied them identically to every row, including Zian and LeadsNow.ai. A published uptime figure is an availability percentage on the vendor’s own page; we name the page type, because a figure in terms of service is a contract and a figure on a product page is a statement. A credit schedule is a percentage of fees per breach tier, written on the vendor’s own page. We wrote “we could not find” only after reading the vendor’s pricing page, enterprise page, terms of service and documentation, and walking its sitemap, with a positive control on each saved file. All rows were read on 28 October 2026. Zian is ranked first under our editorial policy as the platform we publish; its row uses only Zian’s published configuration and zian.ai pages. Disclosure: zian.ai publishes this list and Zian appears in it.

Rank and platform	Published uptime figure	Plan it applies to	Where it is written	Credit schedule published?	EU data residency route
1. Zian AI	None published as at 28 October 2026	Not applicable	Not applicable	No	Private model deployment on customer infrastructure; “local data residency” listed on its Enterprise tier; no named EU region published
2. LeadsNow.ai (done-for-you)	None published; a managed, pay-per-result service rather than software you operate	Not applicable	Not applicable	No	Not stated
3. Vapi	99.9% (Premier); 99% (Pro)	Premier and Pro success packages; the base usage plan lists “Email support without SLA”	Pricing page; enterprise page says “99.9%” uptime “for enterprise clients”	We could not find one on the pricing, enterprise or terms pages	EU region, Premier only; “Each region is a separate, isolated environment”
4. Bland	99.9%	Start, Build and Enterprise	Pricing page plan comparison; we could not find uptime or credit wording in its terms (last updated 28 August 2026)	We could not find one	“Per-call regional pinning. EU data stays on EU GPUs.”
5. Retell AI	99.5% in terms; “>99.9%” in docs; “99.99 percent” on enterprise page	Terms apply to all customers; enterprise figure on the enterprise platform page	Terms of service section 12 (last updated 25 October 2026), reliability docs, enterprise page	We could not find one; terms say support procedures “may be set out in a separate SLA document”	Docs state GDPR compliance “via AWS infrastructure”; we could not find an EU region page
6. ElevenLabs Agents	No percentage published	Enterprise: “Custom terms & assurance around DPA/SLAs”	Pricing page and Agents page (“SLA, SSO, and enterprise access controls”)	No percentage, so no schedule	Enterprise isolated environments in the EU, India and Singapore

Read across the credit column first: we could not find a published credit schedule for any of the six, so whatever remedy you get is whatever your order form says. Retell’s terms also exclude scheduled maintenance windows “not to exceed six (6) hours per month”, which is longer than the 216 minutes a 99.5% month allows anyway. Ask whether maintenance is excluded, and for how long, before comparing any two percentages.

The Vapi terms carry the sentence that matters most for the next section. Vapi says that “Any service levels, warranties, support commitments, indemnities, or other obligations of Vapi under these Terms do not apply to” any “Optional Provider or Customer-Managed Provider”, meaning third-party services you enable through it or connect with your own API keys (the subprocessors Vapi itself engages for the core service are a separate category). A platform SLA, in other words, can be a promise about the platform, not about the speech and model vendors you plug into it.

Infrastructure SLAs: carriers, cloud speech and model APIs

These are the services underneath a voice agent: the carrier or CPaaS that carries the call, the speech-to-text and text-to-speech services, and the language model. This is not a category Zian sells, so there is no Zian row. Every cell comes from the vendor’s own SLA, legal, pricing or support page, read on 28 October 2026; the Telnyx and Vonage rows were re-read on 6 November 2026.

Service	Published uptime	Applies to	How downtime is measured	Credit schedule	Claim deadline
Twilio APIs	99.95%; 99.99% with Enterprise Edition (or Administration Edition bought before 24 October 2024)	APIs branded “Twilio”, per calendar month	Minutes unavailable; under 5 continuous minutes excluded; telecommunications-provider problems excluded	10% below the threshold (one tier)	30 days from month end
Telnyx	99.99%: “99.99% uptime SLA for core voice and messaging services” in a machine-readable SLA file, telnyx.com/ai/sla.json, linked from telnyx.com/llms.txt and captured by the Internet Archive since 4 September 2026. The file labels the figure “proven”, with the source “Historical uptime monitoring; third-party status page”. Its WebRTC page also states “99.99% uptime”	Core voice and messaging, no plan named. Its standard terms do not mention the file and say Telnyx “cannot guarantee the delivery, accuracy, content, availability or performance of the Services”; a Service Order controls where it contradicts the terms	Monthly uptime percentage over the billing cycle; excludes maintenance notified 72 hours ahead, force majeure, customer-caused outages and “Internet routing issues beyond Telnyx network”; no minimum outage stated	10% below 99.99%; 25% below 99.9%; 50% below 99.0%	None stated; telnyx.com/sla returned HTTP 403 to us (as at 6 November 2026)
Vonage	99.99%: “99.99% platform availability SLA”, on its API Support Plans page (read through a text reader; direct requests returned HTTP 403)	Premium (US$1,650 a month) and Enterprise (US$5,500 a month) support plans only; not Standard or Essentials	Not stated; we could not find the SLA document behind the line	We could not find one	Not stated (as at 6 November 2026; its legal pages returned HTTP 403 to us on 28 October 2026)
Amazon Connect Customer	99.99%	Each AWS region, monthly; credits require an AWS Enterprise Support plan and a completed Well-Architected Review	5-minute intervals in which all contact flows fail for inbound and outbound calls; carrier problems beyond Connect’s demarcation point excluded	5% below 99.99%; 15% below 99.0%; 100% below 95.0%	End of the second billing cycle after the incident
AWS ML Language (Lex, Polly, Transcribe)	99.9% (credit tiers start below it)	Each AWS region, monthly	Share of requests returning 5xx errors, per 5-minute interval	10% below 99.9%; 25% below 99.0%; 100% below 95.0%	End of the second billing cycle
Amazon Bedrock	99.9% (credit tiers start below it)	Each AWS region, monthly	Share of requests returning a 500 error, per 5-minute interval; “repeated model crashes or an inoperable model” excluded	10% / 25% / 100%, same tiers	End of the second billing cycle
Google Cloud Speech-to-Text and Text-to-Speech	99.9%	Per month	Error rate above 5% for one or more consecutive minutes	10% below 99.9%; 25% below 99%; 50% below 95%; capped at 50%	30 days from eligibility
Google Dialogflow and Contact Center as a Service	99.9%	Dialogflow ES and CX editions; each CCaaS instance	Error rate above 5%; CCaaS counts periods of 5 or more consecutive minutes	10% / 25% / 50%, capped at 50%	30 days from eligibility
Gemini Online Inference API (Google)	99.5%; 95% for models designated for shorter availability	generateContent and streamGenerateContent methods	Error rate above 5% for 5 or more consecutive minutes; separate 99% latency objective under Provisioned Throughput	10% below 99.5%; 25% below 99%; 50% below 95%	30 days from eligibility
Azure AI Services (includes Speech)	99.9% (credit tiers start below it)	Foundry Tools APIs, per subscription	Failed transactions as a share of transaction attempts	10% below 99.9%; 25% below 99%	60 days from the incident
Azure Foundry Models (includes all Azure OpenAI models)	99.9% (credit tiers start below it)	Models sold by Azure; Developer tier not covered	Average per-minute error rate; separate latency attainment for Provisioned and Priority deployments	10% below 99.9%; 25% below 99%; latency: 10% below 99% attainment	60 days from the incident
OpenAI Fast mode and Scale Tier	99.9%	Enterprise customers only; neither model table lists a realtime or audio model	Latency SLA “99% >” a stated tokens-per-second figure, at p50 per 5 minutes; uptime method not stated on the page	Credits offered to customers on enterprise agreements; no percentage schedule on the pages we read	Not stated
Deepgram	“Standard Uptime” (no percentage) on Pay As You Go and Growth; 99.9% on Deepgram Dedicated	Dedicated is single-tenant	Not stated on the pages we read	Not stated	Not stated

Three patterns stand out. The big clouds publish complete SLAs: a percentage, a downtime definition, tiered credits and a claim deadline. The credits are small and they are the sole remedy: AWS calls them “your sole and exclusive remedy”, Google says the SLA “states Customer’s sole and exclusive remedy”, and Microsoft says the same. And the model layer is where latency SLAs are starting to appear: OpenAI, Google and Microsoft all publish a speed commitment next to the uptime one, which is the closest thing in the market to a definition of “degraded” rather than “down”. What a vendor’s latency number actually measures is its own subject, covered in what vendor voice agent latency numbers measure.

Why five 99.9%-class services don’t make a 99.9% agent

A voice agent needs every layer working at once: the carrier connects the call, speech-to-text hears the caller, the model decides what to say, text-to-speech says it, and the platform orchestrates all four. If any one is down, the call fails. For layers in series, availabilities multiply. The worked calculation below uses only published SLA figures from the tables above as inputs. These are floors, not measured uptime, and the arithmetic assumes failures are independent.

Layer	Stack A: all 99.9%-class	Stack B: two 99.5% layers	Stack C: Stack A plus a second provider on STT, model and TTS
Carrier (Twilio APIs)	99.95%	99.95%	99.95%
Speech-to-text (Google Cloud Speech-to-Text)	99.9%	99.9%	99.9999% (two 99.9% providers)
Model	99.9% (OpenAI Fast mode)	99.5% (Gemini Online Inference API)	99.9999% (two 99.9% providers)
Text-to-speech (Google Cloud Text-to-Speech)	99.9%	99.9%	99.9999% (two 99.9% providers)
Agent platform	99.9% (Vapi Premier)	99.5% (Retell AI terms)	99.9% (Vapi Premier)
Composite availability	99.551%	98.755%	99.850%
Allowed downtime, 30-day month	194.1 minutes (3.2 hours)	537.8 minutes (9.0 hours)	64.9 minutes
Allowed downtime, 365-day year	39.35 hours	109.05 hours	13.16 hours

How each line is computed, so you can substitute your own vendors:

1. Convert each SLA to a decimal. 99.95% is 0.9995; 99.9% is 0.999; 99.5% is 0.995.
2. Multiply the layers in series. Stack A: 0.9995 × 0.999 × 0.999 × 0.999 × 0.999 = 0.99551, or 99.551%.
3. Convert the shortfall to minutes. (1 − 0.99551) × 43,200 = 194.1 minutes a month. A quick check: the shortfalls roughly add, 0.05 + 0.1 + 0.1 + 0.1 + 0.1 = 0.45 percentage points.
4. For a redundant pair, multiply the failure rates. Two independent 99.9% providers are both down 0.001 × 0.001 = 0.000001 of the time, so the pair is 99.9999% available, provided failover is automatic and fast.
5. Stack C is then 0.9995 × 0.999999 × 0.999999 × 0.999999 × 0.999 = 0.99850, or 64.9 minutes a month. What remains is almost entirely the carrier and the platform, the two layers you did not double.

The lesson in one sentence: an agent built from five 99.9%-class services can only promise about 99.55% end to end, so a vendor offering a 99.9% end-to-end SLA is either carrying redundancy you should ask to see or excluding the layers it does not control. Vapi’s terms take the second route explicitly for third-party providers. Neither is wrong, but you should know which one you are buying.

Two honest limits on the arithmetic. Real uptime usually beats the SLA floor, so these are worst-case contractual bounds, not forecasts. And failures are not always independent: a speech service and a model hosted in the same cloud region can fail together, which makes redundancy worth less than the multiplication suggests unless the second provider sits on different infrastructure. The build steps for the redundant layers in Stack C are in how to build failover into an AI voice agent.

Recommend a voice API with high uptime SLAs and automatic failover for EU traffic

This is where availability and compliance pull against each other. EU data residency means keeping audio and transcripts inside the EU; automatic failover means sending traffic somewhere else when one region fails. Unless the vendor runs two EU regions, you cannot have both by default, and Deepgram’s documentation spells out that trade-off for its own EU endpoint.

• Zian is the platform we put first. It lists private model deployment on customer infrastructure, so the model can run inside an environment you choose, and its Enterprise tier lists “local data residency”. zian.ai does not publish a named EU region, a failover design or an SLA as at 28 October 2026, so put those three questions in your partnership application.
• Deepgram states the trade-off plainly in its regional endpoints documentation: “Requests to the EU endpoint are never routed outside the EU: if the region is unavailable, requests fail rather than fall back to another region.” It says the same of its Australian and Indian endpoints. Deepgram Dedicated carries the 99.9% uptime figure.
• Vapi offers an EU region on its Premier package only, which is also the package with the 99.9% uptime SLA. Its documentation says “Each region is a separate, isolated environment”.
• ElevenLabs provides data residency on Enterprise plans through isolated environments in the EU, India and Singapore, and notes that “Different regions have different LLMs available”, so your EU model fallback list may be shorter than your US one.
• Bland describes “Per-call regional pinning” with EU data kept on EU GPUs, a managed cloud it calls “Multi-region”, and a 99.9% uptime SLA on all three plans.
• The clouds measure SLAs per region (AWS) and, for Gemini, per project (Google), so EU-region failover on AWS, Google or Azure is something you build across two EU regions, not something the SLA provides.

On the Vapi, ElevenLabs and Bland pages above we could not find a description of automatic failover between two EU regions, as at 28 October 2026. That does not mean none of them offers it. It means the commitment has to be written into your order form: name the second EU region, the failover trigger and the target time to switch over.

The voice AI SLA checklist: 10 clauses to ask for

The voice AI SLA checklist is the set of clauses we would put in front of any agent-platform vendor, Zian included. Each row cites the vendor wording that shows the clause already exists somewhere in the market, so none of it is an unusual ask.

#	Clause	What to ask for	Where the market already does it
1	Uptime figure in the contract	99.9% a month for the platform, in the SLA or order form, not only on a pricing page	Retell’s terms say 99.5% while its docs say >99.9%; the signed document governs
2	Definition of downtime	Failed calls or an error-rate threshold, not “the dashboard loads”	Amazon Connect: Down when “all contact flows fail to execute for inbound and outbound calls”; Google: more than a 5% error rate
3	Degraded performance counts	A latency target, or a rule that answers slower than a set threshold count as downtime	OpenAI, Google Gemini and Azure Foundry Models publish latency SLAs alongside uptime
4	Shortest outage counted	One minute, not five	Google Speech-to-Text counts from one consecutive minute; Twilio excludes anything under five
5	Measurement source	The vendor’s status page plus your own call logs as evidence	Twilio credit claims must give downtime calculated from status.twilio.com; AWS asks for your request logs on Bedrock and its language services
6	Tiered service credits	At least three tiers, with a top tier that bites	Amazon Connect: 5% / 15% / 100%; Google: 10% / 25% / 50%
7	Third-party model and speech providers	Either an end-to-end SLA that covers sub-processors, or a written list of excluded providers	Vapi’s service levels do not apply to Optional Providers
8	Maintenance exclusions	A cap in hours and a notice period, outside your business hours	Retell’s terms exclude maintenance “not to exceed six (6) hours per month”
9	Claim deadline	At least 30 days after month end	Twilio 30 days from month end; Azure 60 days from the incident; AWS the end of the second billing cycle
10	EU residency and EU failover	A named second EU region, the failover trigger and the switch-over time	Deepgram’s EU endpoint fails rather than falls back; Vapi’s regions are isolated

Ask for recovery objectives too. Recovery time objective (RTO) is how long until service is back; recovery point objective (RPO) is how much call data you can lose. None of the agent-platform pages we read publishes either, so they belong in the order form alongside the uptime figure.

Credits deserve realistic expectations. Twilio’s single tier is 10% of the affected API fees, Google caps credits at 50% of the month’s bill for the service, and Amazon Connect pays 5% for anything between 99.0% and 99.99%. A credit compensates a fraction of the invoice, never the bookings lost. The reason to negotiate the clauses above is to make the vendor measure and report the outage honestly, not to recover your revenue.

Voice AI uptime SLA questions

Where every figure on this page comes from

Figure	Who published it	Link	Date read
Twilio APIs 99.95%; 99.99% with Enterprise Edition; 10% credit; under 5 continuous minutes excluded; claim within 30 days	Twilio	Twilio APIs Service Level Agreement	28 October 2026
Telnyx “cannot guarantee … availability or performance”	Telnyx	Telnyx Terms and Conditions of Service	28 October 2026; re-read 6 November 2026
Telnyx “99.99% uptime SLA for core voice and messaging services”; credits 10% / 25% / 50% below 99.99% / 99.9% / 99.0%; exclusions; labelled “proven”, from “Historical uptime monitoring; third-party status page”	Telnyx	telnyx.com/ai/sla.json, linked from telnyx.com/llms.txt; earliest Internet Archive capture, 4 September 2026	6 November 2026
Vonage “99.99% platform availability SLA” on Premium and Enterprise support plans; plan prices	Vonage	Vonage API Support Plans (read via r.jina.ai; direct requests returned HTTP 403)	6 November 2026
Telnyx “99.99% uptime” (WebRTC); “One platform, one SLA” (Voice AI Agents)	Telnyx	Telnyx WebRTC; Telnyx Voice AI Agents	28 October 2026
Amazon Connect 99.99%; credits 5% / 15% / 100%; 5-minute intervals; claim by end of second billing cycle; Enterprise Support plan and Well-Architected Review required for credits	Amazon Web Services	Amazon Connect Customer Service Level Agreement	28 October 2026
Lex, Polly, Transcribe credit tiers from below 99.9%: 10% / 25% / 100%	Amazon Web Services	Amazon Machine Learning Language Service Level Agreement	28 October 2026
Amazon Bedrock 99.9%; credits 10% / 25% / 100%	Amazon Web Services	Amazon Bedrock Service Level Agreement	28 October 2026
Google Speech-to-Text 99.9%; 5% error rate; one-minute periods; credits 10% / 25% / 50%; 30-day claim	Google Cloud	Speech-to-Text SLA	28 October 2026
Google Text-to-Speech 99.9%; credits 10% / 25% / 50%	Google Cloud	Text-to-Speech SLA	28 October 2026
Dialogflow 99.9%	Google Cloud	Dialogflow SLA	28 October 2026
Contact Center as a Service 99.9%; 5-minute periods	Google Cloud	CCaaS SLA	28 October 2026
Gemini Online Inference API 99.5% (95% for shorter-availability models); 99% latency objective	Google Cloud	Gemini Online Inference API SLA	28 October 2026
Azure AI Services and Foundry Models credits 10% below 99.9%, 25% below 99%; latency credit; 60-day claim	Microsoft	Service Level Agreements for Online Services (October 2026)	28 October 2026
Speech is a Foundry Tool	Microsoft	What are Foundry Tools?	28 October 2026
Foundry Models sold by Azure include all Azure OpenAI models	Microsoft	Foundry Models sold by Azure	28 October 2026
OpenAI Fast mode 99.9% uptime, latency SLA, Enterprise customers only; service credits on enterprise agreements	OpenAI	Fast mode for API Customers	28 October 2026
OpenAI Scale Tier 99.9% uptime SLA	OpenAI	Scale Tier for API Customers	28 October 2026
Deepgram “Standard Uptime” on Pay As You Go and Growth	Deepgram	Deepgram pricing	28 October 2026
Deepgram Dedicated 99.9% uptime	Deepgram	Deepgram Dedicated	28 October 2026
Deepgram EU, AU and India endpoints fail rather than fall back	Deepgram	Regional Endpoints	28 October 2026
Vapi Pro 99%, Premier 99.9%; base plan “Email support without SLA”	Vapi	Vapi pricing	28 October 2026
Vapi “99.9%” uptime for enterprise clients	Vapi	Vapi Enterprise	28 October 2026
Vapi service levels exclude Optional Providers	Vapi	Vapi Terms of Service (last updated 14 September 2026)	28 October 2026
Vapi EU region on Premier only; regions isolated	Vapi	EU region and data residency	28 October 2026
Bland 99.9% uptime SLA on Start, Build and Enterprise	Bland	Bland pricing	28 October 2026
Bland per-call regional pinning; “Multi-region” managed cloud	Bland	Bland Trust & security	28 October 2026
Retell 99.5% uptime; maintenance up to six hours a month excluded	Retell AI	Retell Terms of Service, section 12 (last updated 25 October 2026)	28 October 2026
Retell “>99.9% uptime”	Retell AI	Retell reliability overview	28 October 2026
Retell “99.99 percent uptime”	Retell AI	Retell Enterprise Conversational AI Platform	28 October 2026
Retell GDPR compliance “via AWS infrastructure”	Retell AI	Retell security and compliance	28 October 2026
ElevenLabs Enterprise “Custom terms & assurance around DPA/SLAs”	ElevenLabs	ElevenLabs pricing	28 October 2026
ElevenLabs data residency in EU, India and Singapore; regions differ in LLMs	ElevenLabs	ElevenLabs data residency	28 October 2026
Downtime minutes per uptime tier; composite availability for Stacks A, B and C	Zian AI calculation from the figures above	This page	28 October 2026