What is cloud telephony? VoIP, PBX and CPaaS explained

Cloud telephony means the routing, numbers, menus, voicemail and recording that used to run on a box in your building now run on a provider's servers and reach you over the internet. The consequences that matter are practical: numbers in countries where you have no office, capacity that changes in minutes, and a phone system other software can talk to. The consequence that surprises people is that whether your call gets answered is now decided partly by an authentication signature and a reputation score attached to the number, before the handset rings.

What cloud telephony actually moves off your premises

Four things move, and it helps to name them because vendors bundle them differently.

The switch. The logic that decides which extension rings, when a call rolls over, and where an unanswered call ends up. This used to be a physical PBX in a cupboard.

The numbers. Direct inward dial numbers become entries in a provider's inventory rather than lines terminating on your premises. This is what makes a London number for a Bengaluru team a five-minute task rather than a project.

The media path. Audio is carried as IP packets across networks you do not own, which is the source of most quality complaints and almost none of them are the provider's fault.

The control surface. Configuration becomes an API and a web console instead of a handset programming sequence, which is the change that lets a CRM place a call.

What does not move is the public telephone network. Somewhere in every call to an ordinary phone number there is still a carrier interconnect, a per-minute cost, and a regulator. Cloud telephony changes who operates the equipment; it does not exempt you from telephony.

The five acronyms, in the order they make sense

They are usually presented as a list of alternatives. They are actually a stack, and reading them in dependency order removes most of the confusion.

VoIP — Voice over Internet Protocol. The underlying method: audio digitised, split into packets, carried over IP. Everything below is built on it. VoIP is a technique, not a product, and a vendor whose differentiator is "we use VoIP" has told you nothing.

Hosted PBX — a private branch exchange, the thing that routes calls inside an organisation, running on a provider's infrastructure. Extensions, transfers, hunt groups, voicemail. This is what most small businesses mean by a cloud phone system.

UCaaS — Unified Communications as a Service. Hosted PBX plus messaging, video and presence in one product. The Microsoft Teams and Zoom end of the market, where voice is one mode among several.

CCaaS — Contact Centre as a Service. Built for queues rather than extensions: skills-based routing, wait-time management, agent occupancy, supervisor monitoring, quality management. What you buy when volume is inbound and high enough that queue mathematics is the problem.

CPaaS — Communications Platform as a Service. Not a phone system at all, but APIs — Telnyx, Twilio, Plivo, Bandwidth and similar — that let software place calls, send messages and manage numbers programmatically. This is the layer most modern sales tools sit on.

The last one is where people get caught out. When a CRM says calling is built in, it almost always means it is built on CPaaS. The question that follows is the one worth asking in a demo: whose carrier account is it? That single answer determines your per-minute cost, who owns your numbers, and what happens to both if you leave.

What actually happens when you place a call, in four steps

Worth understanding because it explains nearly every failure mode you will meet.

1. Signalling. Your client tells the platform to set up a call, normally over the Session Initiation Protocol defined in RFC 3261. This negotiates who is calling whom and which audio codecs both ends support.
2. Routing. The platform decides how to reach the destination — over IP to another endpoint, or handed to a carrier that connects to the public telephone network.
3. Media. Audio flows, typically over the Real-time Transport Protocol defined in RFC 3550, often on a different network path from the signalling.
4. Presentation. The receiving network decides what caller ID to show and, increasingly, whether to label the call as suspected spam.

Steps three and four are where the interesting failures live, and they get a section each.

Why a call can connect and still have no audio

Because signalling and media are separate flows that can succeed and fail independently. The call setup completed — which is why both parties see a connected call — and the audio path did not, which is why nobody can hear anything. That is almost always a network or firewall problem at one end rather than a carrier problem.

Voice is unusual in what it demands from a network. Bandwidth is trivial; a single call needs less than a video thumbnail. What voice cannot tolerate is variation. ITU-T Recommendation G.114 on one-way transmission time states that if delays are kept below 150 ms, most applications, both speech and non-speech, will experience essentially transparent interactivity, and that one-way times of 150 to 400 ms are acceptable provided administrations are aware of the impact on transmission quality. Above 400 ms it describes the delay as unacceptable for general network planning purposes.

That threshold is why "we have plenty of bandwidth" is not an answer to a call-quality complaint. A congested link with ample throughput can still deliver packets with enough jitter to make speech choppy, and a network that prioritises voice traffic with the same total capacity will sound dramatically better. Before escalating quality problems to a provider, check three things at your end: whether voice traffic is prioritised, whether the firewall is doing anything clever to SIP, and whether the affected users are on a VPN that adds a detour.

Presentation: attestation, and why a perfect call gets labelled

This is where most outbound calling problems now live, and it is the part that has changed most in the last five years.

Under the FCC's caller ID authentication framework, the originating provider signs a call and attaches an attestation. A, or full attestation, means the provider can confirm the identity of the caller and their legitimate right to use the calling number. B, or partial, means it can confirm the caller's identity but not the number. C, or gateway, means the provider is the entry point to the IP network for a call that originated elsewhere, such as one arriving from abroad or from a network that does not support the framework. The FCC's order required originating and terminating providers to implement the framework in the IP portions of their networks by 30 June 2021, with an extension to 30 June 2023 for smaller providers of 100,000 or fewer subscriber lines.

Attestation is not the whole decision. Calls are not blocked for being B or C — they simply carry less trust into the analytics engines that terminating carriers use to decide whether to label a call. Those engines also weigh behaviour: a number producing high volumes of very short, unanswered calls looks like a robocaller regardless of how it was signed.

Where this is configured and what each strategy selects is set out on the voice platform page, alongside the number inventory the strategies choose from.

Three things help, in descending order of effect:

• Do not burn numbers. Rotation spreads load across a pool, but it is a mitigation, not a cure for a pattern that should not be running. A labelled number recovers slowly.
• Get the highest attestation your carrier can give you, which usually means the carrier knowing that the number you are presenting is one you actually hold on their platform.
• Call at reasonable hours in the recipient's local time, which is also a legal requirement in most jurisdictions and happens to be the same behaviour that protects reputation.

Autocloz's default caller-ID strategy is country_match, which picks an organisation-owned number in the lead's own country, with agent_did and campaign_did as the alternatives. Getting attestation right is the carrier's side of the same problem, and the mechanics of the calling rules themselves are in what TCPA compliance actually requires.

Emergency calling is the obligation nobody prices

Every cloud telephony evaluation covers features and price. Almost none covers this, and it is a legal obligation rather than a feature.

VoIP handles emergency services differently from a traditional line because the network cannot infer where the caller is. In the United States, FCC rules implementing Kari's Law and Section 506 of the RAY BAUM'S Act require that multi-line telephone systems support direct 911 dialling without needing a prefix, notify a designated on-site contact when a 911 call is made, and convey dispatchable location. The location compliance dates were 6 January 2021 for on-premises fixed devices and 6 January 2022 for on-premises non-fixed devices and off-premises devices, when technically feasible, with manual update or alternative location information where it is not.

The operational consequence for a distributed team is that every user's registered address has to be current, and "current" means updated when someone moves house, not when they joined. Ask a prospective provider three things: how the address is registered, how a user updates it themselves, and what the caller experiences if the address is stale.

What carrier pass-through means for your bill

Voice has a genuine marginal cost — carriers charge per minute for the connection to the telephone network. How a vendor handles that cost is one of the largest and least visible differences between platforms.

Bundled minutes. You pay a seat price including an allowance. Simple and predictable, and expensive at volume, because the bundle is priced for the average customer rather than for you.

Resold minutes. The vendor buys wholesale and sells to you with margin. Usually cheaper than a bundle, and still a markup on a cost that scales directly with how much you call.

Pass-through. You bring your own carrier account and the vendor charges a platform fee. You pay the carrier at whatever rate you negotiated.

Autocloz uses the third model: connect your own Telnyx, DIDLogic or FreJun account and those per-minute rates reach your bill unchanged. At a few thousand dials a month the carrier line typically exceeds the software line, which is why the model matters more than the seat price — and it is also why unlimited users on the paid tiers is a coherent position rather than a giveaway, since the marginal cost of a rep is carried by the carrier relationship you already hold. The trade is real and worth stating: pass-through means you manage a carrier account, a top-up and number provisioning. For a team making a handful of calls a day, a bundle's simplicity is worth more than the saving. The full per-tier picture is on the pricing page.

India: the OSP framework, and why cloud telephony is explicitly allowed

Worth a section because the rules changed recently and a lot of guidance still describes the old regime.

India's Department of Telecommunications governs voice-based and data-based service operations — call centres, BPOs, telemarketing — through the Other Service Provider framework. Guidelines dated 5 November 2020 removed the mandatory registration requirement for OSPs, so establishing an OSP centre in India no longer requires registering with the DoT, while retaining the telecom and data security safeguards. A further revision on 23 June 2021 removed the distinction between domestic and international OSPs. Under the liberalised framework, call centres may use internet-based calling, cloud telephony and hosted EPABX systems.

What did not go away is the prohibition on toll bypass, which the guidelines define as the illegal carriage of voice calls between public networks — at the domestic end in India and a foreign country, or between two cities in India — over the OSP's own network rather than that of an authorised telecommunications service provider. In plain terms: use an authorised carrier for the leg that touches the public network, and do not build a private bridge around it.

This is a summary for planning rather than legal advice, and the framework has been revised twice in recent years, so confirm the current text before designing an architecture around it. The wider set of India-specific considerations for outbound sales is on the India-first page.

Do you need a phone system or a dialler?

Different products for different problems, and buying the wrong one is common enough to be worth a section.

A phone system is for receiving calls: numbers, menus, routing, voicemail, extensions, and a queue when volume justifies one. A dialler is for making them at volume: list management, dial modes, dispositions, connect-rate reporting, and the compliance gates that go with outbound.

Most sales teams need mostly the second with a little of the first — an inbound path so callbacks reach somebody, but nothing resembling a support queue. Teams whose volume is genuinely inbound and queued should be looking at CCaaS instead, and no amount of dialler features substitutes.

One thing worth checking either way: where the outcome lands. If the phone system and the CRM are separate products, the call outcome, the recording and the next step travel across an integration you now own and maintain. Native telephony inside the CRM removes that seam, and the argument for it is not features — it is one fewer place for data to diverge. Autocloz's free plan includes the call channel at 100 dials a day alongside 5 users and 10 mailboxes, so the seam is testable before it is bought — start free. The dial modes themselves are compared in auto dialler against power and predictive dialling, and the platform-by-platform view is in the best call centre CRM comparison.

What cloud telephony does not fix, and what Autocloz does not do

The limits, because a category explainer that ends on advantages has not explained the category.

Cloud telephony does not fix number accuracy, and a wrong number is a dial paid for in full with zero chance of connecting. It does not fix labelling once a number has been burned; reputation recovers on the carrier's schedule, not yours. It does not remove the per-minute cost — it only changes who invoices you for it. And it does not improve a network you do not control, which is where most quality complaints originate.

And Autocloz specifically, because the capability differences between carriers are real and we publish them rather than smoothing them over. Not every provider can run every dial mode, and the product refuses rather than silently downgrading: parallel dialling requires the carrier to support outbound dialling, answering-machine detection, server-side leg bridging and simultaneous legs per agent, which among the supported carriers is Telnyx only. FreJun and DIDLogic are rep-console providers — the agent's browser or softphone places the call, so a supervisor process cannot originate calls server-side — and DIDLogic is SIP-based with no carrier-side answering-machine detection at all. Outbound IVR needs both a REST dial API and a programmable answer webhook, which again is Telnyx only today; FreJun's IVR builder is dashboard-configured and inbound-only. Where a provider does not document a parameter, Autocloz sends none rather than an invented one, because a made-up key that a carrier silently ignores bills you for a feature you never receive. Voicemail drops are a management surface — you can register audio and bind it to a scope, but no carrier integration currently plays that audio into a mailbox, so the rep leaves the message. Recording law varies significantly, including between US states, and the platform enforces the retention policy you configure rather than knowing which consent rule applies to a given call. Which carrier you connect is documented per provider, starting with the Telnyx integration.