Salesforce Acquires Cresta: What This Means for AI Agents in Customer Service

A Market That Is Gaining Ground

For Salesforce, Cresta is the latest in a series of acquisitions in the customer experience space that, in 2026, includes, among others, Contentful, Fin (formerly Intercom), Listen Labs, and Braze. According to CMSWire estimates, total spending would exceed $15 billion in a single year. But the trend extends beyond any single company: the major players in customer service software are acquiring startups and competitors at a rapid pace, and the market is consolidating around a few large ecosystems.

The significance of this trend is clear. AI agents for customer service are moving beyond the pilot phase and into full-scale deployment. The debate is no longer about whether to automate customer conversations, but how to do so at scale, with what safeguards, and using what architecture. According to a Gartner forecast from March 2025, by 2029, Agentic AI will autonomously resolve 80% of the most common customer service requests, resulting in a 30% reduction in operating costs.

For those of us who have been developing and managing Agent AI in production for some time, this confirms that we are on the right track. It’s also the right time to clarify how we view this phase and where we stand.

Why the voice, of all things?

Salesforce justified the acquisition by pointing to the central role of the phone channel—the one where it’s hardest to deliver good service. It’s a valid point. The most complex and urgent requests come in over the phone; customers expect a solution during the call, and there is little room for error. For years, it has also been the channel where automation has yielded the most disappointing results: touch-tone menus, rigid voice recognition, and constant transfers.

Next-generation language models have changed this landscape. Today, a voice AI agent can hold a natural conversation, verify the customer’s identity, review their history, and interact with business systems to resolve the request. Consequently, the way service is measured has also changed: it’s no longer about how many calls are “handled” without reaching an agent, but how many requests are resolved on the first contact.

Our experience, however, adds another dimension. Voice is rarely an isolated channel: customers often call after messaging via chat or receiving an email, and the quality of the response depends on how much the agent knows about those interactions. For this reason, in our approach, voice is not a standalone product, but one of the channels within a unified system that also includes WhatsApp, chat, email, and the web—and extends all the way to the systems where cases are actually closed, from CRM to billing to ERP.

Single platform or open architecture

Salesforce justified the acquisition by pointing to the central role of the phone channel—the one where it’s hardest to deliver good service. It’s a valid point. The most complex and urgent requests come in over the phone; customers expect a solution during the call, and there is little room for error. For years, it has also been the channel where automation has yielded the most disappointing results: touch-tone menus, rigid voice recognition, and constant transfers.

Next-generation language models have changed this landscape. Today, a voice AI agent can hold a natural conversation, verify the customer’s identity, review their history, and interact with business systems to resolve the request. Consequently, the way service is measured has also changed: it’s no longer about how many calls are “handled” without reaching an agent, but how many requests are resolved on the first contact.

Our experience, however, adds another dimension. Voice is rarely an isolated channel: customers often call after messaging via chat or receiving an email, and the quality of the response depends on how much the agent knows about those interactions. For this reason, in our approach, voice is not a standalone product, but one of the channels within a unified system that also includes WhatsApp, chat, email, and the web—and extends all the way to the systems where cases are actually closed, from CRM to billing to ERP.

Per-Workstation Licenses and AI Agents: A Model in Need of Reevaluation

There is also one aspect that almost never appears in acquisition announcements, but that anyone choosing a customer service solution would do well to consider: the pricing model. For over twenty years, CRM and contact center platforms have primarily been purchased on a per-seat basis: more agents, more licenses. It was a sensible unit of measurement as long as the work was tied to the people performing it.

AI agents change this equation. Their value lies precisely in handling growing volumes without having to proportionally expand the team: AI handles repetitive requests, while people focus on complex, high-value cases. If pricing remains tied to the number of workstations, the efficiency generated by AI and the cost of the technology risk moving in different directions. This is what analysts call “seat compression”: In its TMT Predictions 2026, Deloitte notes that AI agents can give a single user the capacity of many, reducing the number of licenses required. And Gartner estimates that by 2030, at least 40% of enterprise software spending will shift toward usage-, agent-, or outcome-based pricing models.

The market is already shifting: leading customer service software providers are supplementing traditional licensing models with per-conversation, pay-as-you-go, or per-resolution pricing options. This is a complex transition for companies that have built their business models around user seats, and it will take time. For those who need to make a decision today, however, the practical question is simple: What does the price I pay depend on? The number of users, the use of the technology, or the results achieved?

The European Issue: Where Is the Data Processed?

There is one issue that often takes a back seat in announcements of major acquisitions, but one that is becoming increasingly important for European companies: where conversations are processed. Every interaction with customer service contains personal and contractual data—and sometimes health or financial information—and every response from an AI agent passes through a model running on infrastructure subject to a specific jurisdiction. In regulated industries, this factor can determine whether a project moves forward or is put on hold.

We’ve made a clear choice on this as well. On October 22, at NVIDIA GTC in Berlin, Polarise launched its own sovereign AI Cloud and announced that Increso is among the first customers to use Polarise Drive, its managed inference platform, in production. Polarise directly owns and manages its own infrastructure and, as stated in the press release, data does not leave the European Union and is not subject to the U.S. CLOUD Act.

“Our customers expect AI agents that are fast, reliable, and fully compliant with European data protection regulations. With Polarise Drive, we use the latest Nemotron models for our agents on European infrastructure, without compromising on either performance or control over our customers’ data. This allows us to bring Agentic AI to customer service much more easily, even in sectors where data is most sensitive.” — Marco Modonesi, CTO of Increso

What to Expect in the Coming Months

The integration of Cresta into Salesforce will begin after the closing, in the first half of 2027. It will be interesting to see how two AI agents—each built on different principles (Fin, developed for digital channels, and Cresta, developed for voice)—will coexist, and whether the commitment to neutrality toward third-party platforms will hold up over time.

For companies, meanwhile, the message is already clear: AI agents have become an integral part of customer operations. The decisions to be made concern architecture, governance, and data control, rather than the technology itself.

“When major market players invest billions in this direction, the path forward is no longer in question. The real question for companies is which model to use to follow that path. We believe in an open approach that builds on the processes and systems customers already have in place and gives them full control over their own data.” — Riccardo Frisieri, Co-Founder of Increso

In Increso we continue to develop Agentic AI for Customer Operations, together with our partners and on European infrastructure.

If you'd like to learn more about these topics or understand how to apply them to your processes, please contact Sara directly on our website.