Marc Benioff and Dario Amodei sat cheery-faced together last week to dispel rumors of the SaaSpocalypse through the announcement of “Claudeforce.” Claudeforce is a strategic arrangement aimed at bringing native Salesforce to Claude (headless Salesforce APIs to allow users to live in Claude as their workspace entirely), Claude to Salesforce (Claude as the default model across Agentforce), and Claude to Slack (Claude as the default model across Slackbot, Claude Bot, etc.).
This arrangement reflects some of the key conflicts we discussed in the last article, “Making Sense: Models, Harnesses, and Infrastructure.” The underlying confession in this Claudeforce announcement is Salesforce accepting a future where they become a headless system of workflows and records, orchestrated through users and, increasingly, through agents that live in Claude. For the first time, they have explicitly stated this in their press release:
“The Claudeforce partnership launches with Salesforce in Claude, a Plugin with 37 prebuilt sales skills that enable sellers and agents to reason over live revenue context, automate pipeline updates, and take governed action right from Claude—with the power of Salesforce directly where sellers work.”
As much as Marc Benioff may scoff in the face of SaaSpocalypse, I’m sure the CEO of the #1 CRM company in the world wouldn’t have enjoyed accepting that the place “where sellers work” is moving from their platform to Anthropic. Another fair question is: why has this taken so long for Salesforce and Anthropic? We are upwards of 3 years into the AI revolution, and the markets already had their “SaaSpocalypse” moment earlier this year, with more than $1 trillion in software market value wiped out and companies such as Salesforce, ServiceNow, HubSpot and Workday falling 20–40% over the course of the year.
Well, the voluminous effort and inertia come from 2 primary aspects: i) the level of effort to make complex, often customized enterprise applications agent-ready, and ii) the implications on mid- to long-term commercial models of seat-based SaaS.
Making Enterprise Software Agent-Ready
At Cybage, we have been building turnkey Salesforce MCP servers for enterprise AI use cases for a while, including for ourselves internally, and anyone who has been working on this knows how stubborn Salesforce APIs can be in surfacing relevant context and how different each Salesforce instance can look in terms of data models and access. To make Claudeforce real, there has been significant effort by Salesforce to build their enterprise harness.
“Salesforce in Claude is made possible by AIforce—Salesforce’s trusted enterprise harness that brings all your business data and workflows to any agent through MCP servers, APIs, and CLI tools, without complicated and costly integrations. For decades, enterprise software required users to manually navigate static UI to get work done. Now that agents can access the data, workflows, and rules directly, software is a system that powers every interface.” (Source)
To become an effective application and workflow layer for Anthropic, there must also be increased standardization of Salesforce instances across enterprises. In Claudeforce, a lot of the customization will happen at the agent layer to define skills, run Claude routines, and modify prompts rather than changes at the software layer, giving governed access to data. In other words, rather than customizing Salesforce dashboards or slicing by new record filters through changes in your Salesforce instance, you would make modifications to the agent skills and prompts, delivering dynamic end-user experiences.
Though there is a large amount of complexity and divergence in how software companies should build enterprise harnesses for agentic use cases, it is an easily accomplishable product and engineering feat for companies of any scale. In reality, headless software is not a new architectural concept. The underlying engineering challenge—decoupling application capabilities from the user interface and exposing them through APIs and services—is well understood. Agentic software primarily raises the bar around discoverability, context, authorization, observability, and governance, rather than requiring an entirely new application architecture. Cybage is working with many of its enterprise software clients on this journey to prepare for an agentic future.
There is another factor here on enterprise readiness to consume their current softwares with this headless mechanism, but this is rapidly evolving as enterprise users become proficient with AI in their day-to-day and will soon actively ask for it.
Monetizing enterprise software application harnesses
The broader commercial implications as leading enterprise softwares become headless harnesses are the harder nut to crack. Firstly, of course, as a leading enterprise software or any ISV in existence, I don’t want to become a headless system of workflows and record without a fight. Headlessness threatens to commoditize much of the application layer, weaken the pricing power behind large enterprise seat-based contracts, and compress margins as customers increasingly question why they should continue paying premium per-user prices for software that is accessed primarily through third-party agents rather than through the vendor’s own interface. Unless you are a headless-native enterprise software business like Twilio or Stripe, where the product was designed from the outset to be consumed programmatically and monetized through usage, transactions, or volume rather than human seats, a shift toward headless consumption can be economically disruptive rather than merely architectural. These businesses work because pricing scales naturally with machine-driven activity and the underlying value delivered; traditional SaaS vendors, by contrast, are being forced to retrofit consumption economics onto businesses whose margins and valuation models were built around high-ARPU, recurring seat licenses.
So my first order of business would be to retain at least some, if not all, of the agentic workflows within the boundaries of my software—similar to what Salesforce is attempting with Agentforce as its native enterprise agentic layer. Cybage is building multiple agentic platforms on Agentforce as the technology of choice.
Today, Salesforce already monetizes this directly: Agentforce Flex Credits cost $500 per 100,000 credits, with a standard Agentforce action consuming 20 credits, or roughly $0.10 per action. The commercial problem is that these economics become much harder to sustain once the agent itself sits outside Salesforce. If Claude is doing the reasoning, orchestration, and user interaction, Salesforce cannot simply charge $0.10 every time Claude invokes a Salesforce capability without creating obvious double monetization: the enterprise is already paying Anthropic for the agent and would then be paying Salesforce again for every agentic action. Salesforce and most enterprise software companies therefore have to decide what they are really monetizing—the agent action itself, or the proprietary workflow behind that action.
This is where the Anthropic partnership becomes strategically important. If Salesforce cannot own every enterprise agent, it can at least try to preserve the commercial bundle around the ones customers choose.
This is important because a fully open, provider-agnostic enterprise harness, while better for enterprise customers and society at large, is strategically dangerous for Salesforce. If Claude, OpenAI, or Gemini can simply plug into Salesforce through a standardized layer of permissions, tools, and workflows, then the architecture becomes cleanly separable: the customer buys intelligence from one provider and treats Salesforce increasingly as a commoditized system of record and execution underneath it. In that world, Salesforce risks losing not only the interface and intelligence layer, but also the pricing power associated with its large integrated enterprise contracts. This would open the ground for disruptive headless open-source systems for enterprise work.
The Anthropic tie-up therefore makes sense not merely as a technical integration, but as a way to keep bundled commercial offerings on the table. Rather than customers independently procuring Claude Enterprise, Salesforce licenses, an agent gateway, and separate workflow capacity, Salesforce and Anthropic can package these layers together—potentially through joint enterprise agreements, bundled seats and consumption, integrated security and governance, and coordinated support. The customer still experiences a single enterprise solution rather than a collection of interchangeable architectural components.
That distinction matters. Salesforce may ultimately expose an open agentic harness to multiple providers, but it is unlikely to want the harness itself to become the product boundary. If every model provider can access Salesforce on identical commercial terms, Salesforce risks being pushed down the stack into a low-differentiation workflow utility. The more attractive strategy is to remain technically open while creating preferentially integrated commercial bundles with leading AI providers, ensuring that Salesforce continues to participate in the economics of the intelligence layer rather than merely supplying the plumbing beneath it.
Implications on other ISVs
Cybage lives and breathes with ISVs. Many of our 250+ customers provide enterprise software at scale across sectors ranging from payments to healthcare to media to retail to supply chain and more. Regardless of industry, all software companies have similar roadmaps of building first-party agentic workloads, rearchitecting legacy applications, and reimagining mid- to long-term pricing models. While not everyone will enjoy the network effects a Salesforce has, many companies occupy hard-to-displace niches in their respective enterprise software domains.
Some software companies, especially those with defensible and differentiated user interfaces, are evolving into offering first-party agentic workflows. Their products will increasingly look like chat interfaces to trigger workflows + centralized agent hubs to monitor agentic actions, escalations, policies, and more. We are accelerating development of these platforms with our learnings from our own agentic platform, CybArgo (more to come on this soon). Other companies will have to accept their long-term position as integration endpoints for AI workspaces they don’t control and differentiate on the quality and design of their enterprise harness, the robustness and reliability of data, and their GTM co-ordination with foundational model labs or other Agentic Platforms. In all cases, there will be disruption of commercial models and innovation needed across product, technology, and sales over the next 5 years.
Authored by
Aneesh Nathani
VP - Data and AI, Cybage Software
