Salesforce Plans to Acquire Cresta: The Next Step for Voice AI Is to Resolve Customer Issues

A customer called to request a change to the order address. The AI understood the request and provided a polite response. However, if the order system wasn’t updated and the warehouse still shipped the order to the old address, the call didn’t actually resolve the issue.

This hypothetical scenario illustrates that the value of voice AI must be validated throughout the business process: recognizing user requests is just the starting point; subsequent steps include identity verification, rule evaluation, system operations, and result confirmation.

On November 2, 2026, Salesforce announced that it had signed a definitive agreement to acquire Cresta. Cresta offers products focused on voice agents, agent augmentation, and conversation analytics. Salesforce plans to integrate these capabilities with Service Cloud to connect customer context, conversations, and service outcomes. The transaction is expected to close in the first quarter of Salesforce’s fiscal year 2028, subject to customary closing conditions, including regulatory approvals; the acquisition should not be considered complete at this time.Official Salesforce Announcement

Why is it important to have customer information and real-time conversations close together?

During a service call, customers usually don’t recount their entire history from the beginning. They might simply say, “That issue from last time still hasn’t been resolved,” or they might switch between two different requests within half a minute.

If a voice system can only process the current statement, it becomes difficult to determine the order status, previous commitments, and available remedies. On the other hand, even if a customer relationship management system maintains a complete record, it still requires an interface capable of understanding the customer’s current request and triggering the appropriate workflow.

This provides the business logic for the integration of these three components: customer data helps determine the context, conversational capabilities help identify real-time needs, and backend processes handle execution. Whether these three elements can be seamlessly integrated determines how much information the customer must repeat, whether a human agent must take over from the beginning, and whether a complete record of the service is maintained.

However, the proposed acquisition is only the starting point for integration. While it’s easy to combine product catalogs, truly integrating permissions, data structures, and exception handling is much more complex.

Cresta’s journey is more inspiring than simply “adding voice capabilities.”

A review of Cresta by Sequoia Capital shows that the company began in 2017 at the Stanford Artificial Intelligence Laboratory, where its early focus was on assisting human agents by providing suggestions during conversations. Later, the product expanded to include AI agents capable of handling entire customer interactions.A Look Back at Cresta by Sequoia Capital

The lesson this approach offers enterprise AI developers is to first understand how a task is actually performed. Customer service involves more than just “answering questions”; it also includes researching information, identifying exceptions, assessing sentiment, confirming commitments, and documenting results. Only by focusing on these details over the long term can developers build up reusable business capabilities.

Therefore, when evaluating a voice agent, in addition to assessing whether it speaks naturally, one should also observe whether it asks follow-up questions when faced with ambiguous requests, whether it stops when encountering requests beyond its scope, and whether it retains sufficient context after handing the matter over to a human agent.

It is difficult to assess these capabilities based solely on a smooth demonstration. The more critical the business process, the more testing needs to account for scenarios that do not go as planned.

The biggest integration challenge is often hidden in the “what's next?”

Take an address change as an example: a seemingly simple request may trigger multiple checks: Does the caller have the authority to modify the order? Has the package already been shipped? Will the address change result in additional shipping fees? If a conflict arises, should the request be transferred to a customer service representative or put on hold?

This means that companies need to clearly define the scope of actions for agents. Authorization for queries and modifications should be separate; operations involving fees or significant commitments should require approval; and there should be a clear recovery process in place for system write failures. Just because the AI has verbally agreed to a customer’s request does not mean that a backend operation that has not yet been completed should be marked as successful.

Manual handoffs are just as important. Customers should not have to explain the entire situation again after being transferred, and the agent taking over needs to know what the previous agent has already looked into, what promises were made, which actions were successful, and which remain incomplete.

From this perspective, competition in voice AI will extend to process design and organizational collaboration. While the model is one component, service reliability also depends on permissions, tool integration, business rules, and accountability.

Don't just focus on the call duration—ask first if the issue has been resolved.

If a company rewards agents solely for shortening call duration, agents may end calls prematurely; if the company focuses solely on automation rates, it may leave complex issues that should be handled by humans to be resolved by machines.

A more reasonable evaluation should consider both efficiency and outcomes. It is possible to track separately whether customers call back, whether requests are properly fulfilled, whether manual handoffs proceed smoothly, and the rework caused by errors. By combining this data with actual operating costs, one can determine whether a particular type of task is suitable for expanded automation.

For customer service managers, it is advisable to start with processes that have clearly defined boundaries and verifiable outcomes, while continuing to rely on human intervention for complex disputes and high-risk commitments. Only when exceptional scenarios can be handled consistently will increasing the level of automation become more meaningful.

The transaction has yet to be finalized; the focus should return to implementation.

For existing customers, the key areas to watch moving forward include the post-merger product roadmap, system compatibility, changes to contracts and pricing, and whether the transition will impact currently running services. None of these should be assumed based solely on the proposed acquisition. According to Salesforce’s announcement, Cresta will continue to support the primary contact center and phone platforms that customers currently rely on after the transaction is completed, but the specific integration plan will depend on future arrangements.Salesforce's Explanation of Transaction and Platform Support

The value of this proposed transaction lies in the fact that it places voice interfaces and enterprise business systems on the same roadmap. How far enterprise AI can go next depends on its ability to transform a single conversation into verifiable service outcomes within a clear framework of permissions and processes. Ultimately, what matters most to customers is whether the task has been successfully completed.