The demos were real. The use cases were compelling. But the leap from "Claude can do this" to "enterprises will rip out Salesforce to get it" relies on a misreading of how corporate technology actually works.
Anthropic's enterprise agent event was designed to show what Claude can do inside a real business. It succeeded on its own terms. Spotify used it to cut engineering time on complex code migrations. Novo Nordisk improved clinical study documentation timelines. Salesforce reduced turnaround times in Slack. For a foundation model company making its case to enterprise buyers, these are exactly the right stories to tell.
Wedbush's analysts watched the same demos and drew a sharply different conclusion from the one the AI disruption trade is built on. Their core argument is that the market is confusing model capability with enterprise software replacement, and that conflation is driving a software sell-off that the underlying business reality does not support.
The distinction matters
What Anthropic and OpenAI demonstrate in product events is raw model intelligence, which is impressive, but not the same thing as the compliance infrastructure, audit trails, security controls, uptime guarantees, billing systems and workflow orchestration that enterprise software platforms have spent two decades building. Microsoft, Salesforce, ServiceNow and Pega are not primarily in the business of being smart.
They are systems of record embedded so deeply in mission-critical operations that replacing them would require dismantling infrastructure that most enterprises have spent a decade integrating. An LLM does not change that calculus. It sits inside the orchestration layer, not above it.
The cybersecurity argument runs in the same direction
AI agents expand the attack surface dramatically, multiplying API connections, machine identities and cloud-native workloads. Every new autonomous workflow is a new vector. Wedbush argues this makes the major cybersecurity platforms, CrowdStrike, Palo Alto Networks and Zscaler, among them, structural winners in an AI world rather than the collateral damage the market has occasionally treated them as.
The deepest point in the Wedbush note concerns distribution. Anthropic and OpenAI do not have 20 years of CIO relationships, vertical-specific workflow integrations, or embedded enterprise sales infrastructure. The model layer, Wedbush contends, will commoditise faster than the application layer, where business logic actually lives.
If anything, the urgency AI is creating around workflow automation is accelerating deal cycles for the platforms that already own those workflows, not bypassing them.
The demos were good. The threat to enterprise software incumbents remains, in Wedbush's view, substantially overstated.