The Salesforce and Anthropic partnership is being framed as another front in the debate over whether AI will kill SaaS. That's the wrong debate.
What Salesforce is doing with Anthropic's Claude is a direct signal of where enterprise software is headed. Rather than simply bolting AI onto an existing interface, Salesforce is opening its underlying data and capabilities to frontier models, effectively separating the intelligence layer from the application itself.
This is the model smart SaaS companies will follow in the future, another step toward an agentic world where AI operates software instead of humans.
Systems of record aren't going away. Salesforce may become even more valuable as the trusted data and business logic layer underneath agents. What's changing is how that information gets put to work.
This is the question for CIOs: What happens to everything that sits between those systems of record and the work itself?
Related:Small language models promise savings. Here's what happens later
Today's enterprise software stack was designed around human operators. Interfaces help people find information, workflow tools route work between them, and layers of software monitor, evaluate and optimize the work. As agents become the primary operators of enterprise workflows, CIOs will need to determine which application layers remain essential when the primary user is no longer human.
Making software operable by agents is only one part of the transition. Giving an agent access to data and tools doesn't give it the domain expertise to put them to work.
General intelligence still needs domain expertise
One detail from the announcement stands out: Salesforce and Anthropic created 37 specialized sales skills for Claude. Why?
General intelligence and domain expertise are very different.
Frontier models reason extraordinarily well. Making that intelligence useful inside a business function requires deeper understanding: what matters, which context to use, what actions are available, how to evaluate success, and how thousands of individual decisions connect to a business outcome.
Architecting that reasoning is no small task. When you give an agent access to your systems, you only give it the tools. You haven't taught it how to run the function. That burden falls on enterprises to define skills, connect context, set guardrails and orchestrate agents. As more expertise moves into agentic systems, the surrounding application stack must change with it.
From AI agents to agentic systems
Agentic AI is beyond individual tasks. It's running entire business functions, and customer experience offers an early blueprint for this model.
Related:Shadow AI has gone agentic. Are organizations ready?
Serving the customer is only one part of running a customer experience operation. You also need to understand why interactions succeed or fail. You must evaluate quality, identify patterns and determine what should change. Then test those changes and make the next interaction better.
Historically, we've built separate applications, or even entire teams, to perform each of those functions. But with AI-native architecture, those functions can become specialized agents.
Specialization alone isn't enough, however. The agents need to be orchestrated: Which agent should reason about a problem? What context does it need? When should another one take over? And how does one agent's output become another's input?
That orchestration is what turns a collection of specialized agents into an agentic system.
How agents are transforming the application stack
Enterprise applications reflect how organizations have historically divided work among people.
A typical enterprise CX stack might include customer service call software, ticketing, workforce management, quality assurance, knowledge management and voice-of-the-customer tools. Each solves a specific problem, but many exist because people need help coordinating customer interactions across fragmented systems.
Related:The 3 organizational barriers blocking AI deployment for CIOs
CIOs should reconsider whether each function requires its own application, given that agents can manage multiple parts of a workflow themselves.
Instead of replacing six applications with six agents, organizations should design the function as a unified agentic system.
The next frontier: Compounding intelligence
Making software accessible to an AI agent isn't the same as designing a system that learns. In an agentic CX system, every interaction can become an input into the next improvement.
The agent serving a customer creates a signal. A quality agent evaluates what happened. An insights agent identifies patterns across interactions. An optimization agent determines what should change. Simulation tests that change before they reach another customer.
Then the next interaction starts with a better system.
The model provides intelligence, and domain specialization turns it into expertise. Orchestration turns that expertise into a system, and the learning loop allows that intelligence to compound.
Advances from companies like Anthropic and OpenAI are good news for vertical AI companies. We don't need to build better foundation models. As those models improve, so do the capabilities of specialized systems.
Instead of trying to predict which SaaS companies survive AI, the focus should be on determining which systems remain essential sources of data, business logic and governance; which application layers agents can absorb; and where rebuilding a function around an agentic architecture creates fundamental operational change.
For CX, a better customer service agent is only the beginning. The real opportunity is a CX operation that learns, adapts and improves with every interaction.