Voice AI consolidation: the four things to keep portable

Between 20 October and 5 November 2026, the voice AI stack consolidated from both ends. Speech-model companies started selling whole voice agents, while contact-centre and CRM suites bought or merged with the companies that run calls. Whatever you buy next, keep four things under your own control: your phone numbers, your agent definition, your call records and your data contract.

Then prove, during the pilot, that you can change one layer without rebuilding the rest. The swap test near the end of this post shows how.

A disclosure: Callaro is an orchestration platform that routes each agent across several model vendors, so we have a side here. We've tried to be specific about when a bundle wins. For the basics of how the layers fit together, see our complete guide to AI voice agents.

What happened in those seventeen days

Two moves ran in parallel, each confirmed by the companies' own announcements.

Model vendors moved up to the whole agent

• 20 October, Mistral: Voxtral 3 paired new transcription and speech models with a voice agent SDK built on LiveKit, with SIP for phone calls.
• 20 October, ElevenLabs: Eleven Realtime, its first speech-to-speech models, compatible with OpenAI's Realtime API, with ElevenAgents as the recommended way to deploy them.
• 20 October, Zoom: Voice Agent APIs that run recognition, reasoning and synthesis as one managed pipeline.
• 27 October, AssemblyAI: a speech-to-text company shipped its first text-to-speech model, AssemblyAI-TTS-1.
• 28 October, Rime: a text-to-speech company shipped Prism, a speech-to-speech model that drafts each reply as text first. The week before, it had added its own speech recognition.
• 29 October, Cartesia: Managed Agents, a hosted end-to-end agent platform with telephony, knowledge bases and tool integrations.
• 29 October, Deepgram: agreed to acquire Daily, the maker of the open-source Pipecat framework. Pipecat keeps its BSD licence, and an independent foundation is planned.

Suites absorbed the agent layer

• 2 November, Salesforce: signed a definitive agreement to acquire Cresta and build its voice agents and conversation intelligence into Service Cloud.
• 4 November, Dialpad and Ada: announced a merger. Their stated reasons include owning Dialpad's voice network and its own speech and language models, for "greater control over core costs and performance".
• 4 November, ElevenLabs: native integrations for ElevenAgents with Genesys Cloud, Amazon Connect, Five9 and NiCE CXone.

There was one counter-current. Orchestration platforms kept adding everyone's models: in the first week of November, Retell made xAI's Grok Voice 3.0 selectable and opened GPT-Live-1 to paid customers.

Why a bundle is tempting, and sometimes right

The strongest argument for buying the whole stack from one vendor is turn-taking. Deciding that the caller has finished is a recognition problem. Starting the reply quickly is a synthesis problem. The gap between the two is where calls start to feel robotic.

A vendor that owns both sides can tune them together. That is Cartesia's case for Managed Agents, and it's a fair engineering argument, though the comparisons behind it are Cartesia's own.

There are practical reasons too:

• One contract and one support line when a call goes wrong at 2 a.m.
• Native placement in your contact centre. If Genesys already carries your calls, an agent that joins through Genesys's own Audio Connector skips a SIP trunking project. One ElevenLabs customer named that integration as an important factor in its choice.
• Native context in your CRM. If your service team lives in Salesforce, a voice agent built into Service Cloud gets customer history without an integration project. Callaro connects to CRMs through signed webhooks and your own APIs rather than a native Salesforce connector, which is more work on day one.

What a bundle costs you later

The bill for a bundle arrives at the first switch.

Leaderboards keep changing hands. On 15 September, Google said Gemini 3.8 Live Extended Thinking ranked first on Artificial Analysis's Speech to Speech Quality Index. On 29 October, Cartesia said Managed Agents ranked first of 30 on Artificial Analysis's Speech Agent Arena. Two boards, six weeks apart, two "number ones". Tie your contract to one vendor's models and you've bought its roadmap, not today's ranking.

Models retire on the vendor's calendar. Google's preview Live models on the Gemini API are scheduled to shut down from 17 November. A platform that treats models as replaceable parts has a path through that; we've written up how a voice AI platform survives model sunsets.

API compatibility isn't behaviour compatibility. Eleven Realtime accepts existing OpenAI Realtime clients after an endpoint and model change. That moves the connection, not the behaviour. When we added a full-duplex speech-to-speech path beside our pipeline in September, only about half of the 105 pipeline behaviours we audited were at parity (what we found).

Cost control moves to the vendor. Dialpad and Ada said it plainly: owning the network and the models gives them control over core costs. That control is theirs, not yours.

On a single bundled usage charge you can't see which layer drives the bill, so you can't trade a cheaper voice for a better recogniser on one call type. A bill split by component (speech recognition, the language model, speech synthesis, telephony) makes that trade possible.

Integrations follow the owner's roadmap. Both deals promise continuity: Dialpad and Ada say existing integrations keep running as configured, and Salesforce says Cresta will keep supporting the contact-centre and telephony platforms its customers use today. Those promises cover today. Where the next connector gets built is the acquirer's call.

The four things to keep portable

1. Your phone numbers

Numbers carry history. Callers recognise them, answer rates attach to them, and in India the ranges that carry automated calls are declared to your operator (our post on TRAI's A2P call rules has the detail). If the agent vendor owns the numbers, leaving it means new numbers and building that history again.

We treat numbers as operating assets, not settings. On Callaro, numbers are bring-your-own by default, and a campaign dials only from numbers on your own telephony connections. Each pool number is weighted by its live answer rate in that campaign, and once a number reaches a contact it sticks for that contact's later attempts. With bring-your-own telephony, your carrier bills you directly.

Ask any vendor three questions. Whose account are the numbers on? Can I port them out? Can I bring my own SIP trunk?

2. Your agent definition

Your agent is more than your prompt. Every platform adds behaviour you didn't write, and that behaviour stays behind when you export the prompt. On our pipeline, for example, the platform:

• speaks the greeting itself and tells the model what the caller has already heard;
• adds voice rules tuned to the chosen speech engine, such as no markdown or lists and how numbers and dates are written for that engine;
• refuses to end a call when the caller spoke last without a closing line;
• holds a transfer the model attempts on the caller's first turn, or with no request from the caller, until the caller confirms.

Paste the same prompt into another vendor's builder and none of that comes with it. Same prompt, different calls.

So keep the source of truth outside any one vendor: the prompt sections, the extraction schema with its field types, the tool definitions (including which parameters the platform fills and which the model decides), and a written list of the platform behaviours you rely on.

Callaro's builder keeps prompts in named sections and snapshots a version on every save and publish. That makes the list easier to maintain. It doesn't make it optional.

3. Your call records

Recordings, transcripts, extracted data and QA scores are the evidence for every decision about the agent, including whether to switch. Two properties matter.

The first is a format another model can re-read. Our recordings are time-aligned stereo, with the caller on one channel and the agent on the other, so either side can be re-transcribed later with a different recogniser.

The second is a scorer that isn't the vendor's own conversation model. If your QA scores come from the same vendor that ran the calls, you can't compare before and after a vendor change on equal terms.

Callaro's post-call analysis prompts run on OpenAI, Gemini or Anthropic models, chosen separately from the model that held the call, and a campaign-wide analysis run snapshots its prompt, schema and model so scores stay comparable.

Ask: can I bulk-export recordings (dual-channel or mixed down?), transcripts with timings and structured outcomes, and for how long after the contract ends?

4. Your data contract

How results reach your systems decides how much of your integration survives a switch. A documented, versioned event contract that you consume is portable. A vendor-built native connector is not, however convenient it is.

Here's what ours carries, as an example of what to look for. Every completed call sends an event with a schema version, an idempotency key that stays the same across retries, the extracted fields as typed values, and separate fields for runtime status, business outcome and telephony status. Deliveries can be signed with HMAC once you set a secret.

A receiver that dedupes on that key and maps typed fields can take another vendor's events through a thin adapter.

Choose models per agent, not per contract

Pick models per agent or per call type, so a better voice for one language doesn't mean a new vendor for every call. The model registry our agent builder reads today lists speech-to-text from six vendors, text-to-speech from seven and two speech-to-speech families, and a retired model maps to its named replacement. The current list is on our models and languages page.

Run the swap test before you sign

A demo can't show you portability. A pilot can, if you plan one swap into it.

1. Pick the layer you're likeliest to change in the next 18 months. For most teams it's the voice for a new language, or the language model.
2. Clone the pilot agent and change only that layer. If the vendor needs a new agent type or a services ticket to do it, you already have your answer.
3. Replay real calls through both agents. Take 20 to 30 pilot calls, stub the tools, and run both agents on the same caller turns. Our guide to testing an AI voice agent shows how to compare what was said, what the tools did, what the caller heard and what happened on the line.
4. Check what should not have moved: the calling numbers, the event payload and its schema version, the extracted field types, and the QA scores from your independent scorer.
5. Time it. Record the hours from decision to a cloned agent passing the replay. That's your switching cost; put it in the vendor comparison next to usage cost.
6. Write the results into the contract: number porting, export formats, how long data stays available after termination, and notice periods for model retirements.

A worked example

A lender runs collections calls in Hindi and English on a bundled stack and wants to try a newer Hindi voice that tested better with its customers.

On a platform that chooses models per agent, the work is the steps above: clone, change the speech model, replay, check, time. On a bundle whose voice is fused with its recognition and turn-taking, there is no separate voice to change. Trying a better Hindi voice becomes evaluating a new vendor and moving numbers, agent, records and integration together. Same goal, two very different projects.

When a bundle is the right call

Buy the bundle, knowingly, when:

• you're standardised on one contact-centre platform and want the agent inside its routing and reporting;
• your service team runs in Salesforce and native context matters more to you than model choice;
• you run one language and one use case at modest volume, so switching costs little either way and speed to launch matters most;
• you don't have the engineering time to own an event receiver.

Even then, keep the numbers on your own carrier account and the records exportable. Both cost nothing to protect at signing and a great deal to recover later.

What this means if you run on Callaro

Nothing in your stack changes because of these deals. Three of the companies above (Cartesia, Deepgram and ElevenLabs) supply models in our registry, and their moves up the stack don't change the components you've selected. If one retires a model, the registry maps agents to its replacement, and the swap test above is how you check the result.

Consolidation won't stop at two deals. Bundles will produce better demos; the buyers who come out ahead will be the ones who can change one layer in a week rather than a quarter. If you'd like to run the swap test on your own calls, we'll set up the pilot with you.