Skip to the stories
Microsoft 365 & SharePoint

Salesforce, Anthropic and the Layers Agents Will Make Optional

  • Andy Lee
  • 370 views

Salesforce's move to open its data to Claude points to agents running workflows. CIOs should check which application layers still earn their place.

Salesforce, Anthropic and the Layers Agents Will Make Optional

The Salesforce and Anthropic tie-up keeps getting pulled into the old argument about whether AI will kill SaaS. Andy Lee, CEO of Crescendo, thinks that's the wrong debate. My read is that the more useful question for CIOs is what happens to everything sitting between the systems of record and the actual work.

The intelligence is separating from the application

Salesforce isn't simply bolting AI onto its existing interface. It's opening its underlying data and capabilities to frontier models, which in effect pulls the intelligence layer away from the application. Lee expects smart SaaS vendors to follow that pattern, and he sees it as another step toward an agentic world where AI operates the software directly.

That doesn't mean systems of record are on their way out. Here's the thing: Salesforce may end up more valuable as the trusted data and business logic layer beneath agents, because that's where the authoritative information lives. What changes is how that information gets put to work.

A stack built for human operators

Most enterprise software assumes a person at the keyboard. Interfaces help people find information, workflow tools pass work from one person to the next, and monitoring and optimization layers watch over the results. When agents become the primary operators of those workflows, a lot of that scaffolding has less to do.

That's why CIOs will need to work out which application layers remain essential once the main user is no longer human. It's a harder exercise than it sounds, and it's one of the reasons agentic governance is starting to look like more than a traditional IT practice.

Tools aren't the same as expertise

Making software operable by agents is only half the job. Giving an agent access to data and tools doesn't teach it how to run a business function. Lee points to one detail from the announcement as telling: Salesforce and Anthropic created 37 specialized sales skills for Claude.

General intelligence and domain expertise are different things. Frontier models reason very well, but turning that reasoning into something useful inside a business function takes more. The agent needs to know what matters, which context to pull in, which actions are available, how to judge whether it succeeded, and how thousands of small decisions add up to a business outcome.

That work lands on the enterprise. Teams have to define the skills, connect the context, set guardrails and orchestrate the agents. As more expertise moves into agentic systems, the application stack around it has to change too. Expect this to be the slow part of most rollouts, since it's mostly process and domain knowledge rather than configuration.

Customer experience as the early blueprint

Lee's argument is that agentic AI is moving beyond individual tasks and toward running entire business functions. Customer experience, in his view, offers the clearest early model.

Serving the customer is only one piece of a CX operation. You also have to understand why interactions succeed or fail, evaluate quality, spot patterns, decide what should change, test those changes and then make the next interaction better. Historically, organizations built separate applications, or even separate teams, for each of those jobs. With AI-native architecture, Lee argues, each function can become a specialized agent.

Specialization alone won't carry the design, though. The agents need orchestration, and that raises practical questions. Which agent should reason about a given problem? What context does it need? When should another agent take over? How does one agent's output become the next agent's input? Answering those is what turns a collection of specialized agents into an agentic system.

From a pile of applications to one system

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 one solves a specific problem, but many of them exist mainly because people need help coordinating customer interactions across fragmented systems.

That's the reason CIOs should question whether every function still needs its own application. Agents can manage several parts of a workflow on their own. Lee's suggestion is blunt: rather than swapping six applications for six agents, design the function as a single unified agentic system.

The payoff is the learning loop. Making software reachable by an agent is not the same as building a system that improves. In an agentic CX setup, the serving agent creates a signal, a quality agent evaluates what happened, an insights agent finds patterns across interactions, and an optimization agent proposes what should change. Simulation tests those changes before they reach another customer. Then the next interaction starts from a better system.

Where the value compounds

Lee frames the stack as a sequence. The model supplies intelligence. Domain specialization turns that intelligence into expertise. Orchestration turns expertise into a system. The learning loop lets the whole thing compound over time.

He also argues that advances from companies like Anthropic and OpenAI are good news for vertical AI companies. Those firms don't need to build their own foundation models, because as the underlying models improve, the specialized systems built on them improve too.

For SaaS vendors and the people who buy from them, Lee's advice is to stop guessing which companies will survive AI and ask more precise questions. Which systems remain essential sources of data, business logic and governance? Which application layers can agents absorb? Where does rebuilding a function around an agentic architecture produce a real operational change rather than a cosmetic one? For CX teams, a better customer service agent is only the starting point. The bigger opportunity is a customer experience operation that learns, adapts and gets better with every interaction.

Lee is CEO of Crescendo, an AI-native customer experience platform. He draws on 27 years of hands-on CX operations experience and founded Alorica in 1999 with $10,000. By his account, that company grew into a business of roughly $2 billion with more than 100,000 employees across 17 countries.

Source: Andy Lee · www.informationweek.com

Neighbouring cells