Why Salesforce bought Cresta after it already built Agentforce Voice
Salesforce built its own voice agent, Agentforce Voice, and showed it on the Dreamforce stage in October 2025. It became generally available this June. Salesforce's own customer stories still don't name a single large company running it in production. Now Salesforce is buying Cresta, a voice AI company, for a reported $2.7 billion. Salesforce — a company with the resources to build almost anything — built voice agents, shipped them, and then decided to acquire one anyway.
Why? Kishan Chetan, who runs Agentforce and Service Cloud, said in the announcement that voice is where customer service is hardest to get right and where it matters most. He's right.
Anyone can wire a couple of speech models to an LLM and have a demo by lunch. It'll sound great. Then real callers show up. They interrupt, wander off topic, and read you a card number mid-sentence. When your agent says the wrong thing, nobody cares that it sounded natural. Confidently wrong is still wrong. And you don't fix that with a better prompt. You fix it by putting the business logic in code, where the model can't talk its way around it.
Every week we meet companies replacing prompt-and-pray voice agents, both commercial products and ones they built themselves. They're switching for the same reason: the agents were unreliable narrators.
Enterprises want control. They want agents that stay on task and follow the business's process, not the model's best guess at it. They need to handle payments and health information without that data ever reaching the model. They need their own compliance controls enforced by the system, not by an LLM that's been asked nicely to behave.
Salesforce's announcement correctly points out that customer conversations get stuck in silos, cut off from customer context and the tools teams use. For voice AI, that's about more than pulling customer data from the CRM. Most voice agents sit beside the phone system rather than in it. The call comes in through a carrier, crosses into a contact center platform, and gets handed to the AI over SIP or a forked audio stream. Every hop adds delay the caller hears as an awkward pause. Every hop is another place for the call to drop. And every hop is a place your code loses control.
Production voice AI is a telephony problem
The hard moments happen in that plumbing. When a caller needs a person, that transfer has to carry the context so they don't start over. When a caller pays by phone, the card digits have to be caught in the call, before the model ever hears them. The AI provider goes down, and calls have to keep flowing. None of that is a speech or LLM problem. It's a telephony problem, and companies that treat voice AI as an app bolted onto someone else's phone system find that out in production.
The market is consolidating around this from both ends. NICE bought Cognigy last year. Deepgram, which built its business on speech models, just agreed to buy Daily, the company behind the open-source Pipecat framework. Now Salesforce is buying Cresta. Others are partnering instead of acquiring. The direction is the same either way: platforms want production-grade voice, and they've decided it's faster to get it from someone who has already done the hard part.
How to evaluate a voice AI vendor
If you're evaluating voice AI, skip the question of whether it sounds natural. Ask the vendor:
- to show you that the model never sees sensitive data
to show you that the model never sees sensitive data
- to show you where the guardrails live. If the answer is a paragraph in the prompt telling the model to police itself, you have your answer
to show you where the guardrails live. If the answer is a paragraph in the prompt telling the model to police itself, you have your answer
- what happens to the call, and to the caller's context, when it needs to go to a person or when a provider has an outage
what happens to the call, and to the caller's context, when it needs to go to a person or when a provider has an outage
The demos got everyone's attention. The acquisitions say the market has moved on to the harder question: running voice AI in production, with real customers, real data, and real rules.
If you're a Cresta customer or partner and you're wondering what independence looks like after the deal closes, we'd like to talk.
