The numbers hit my terminal like a shockwave. Salesforce, the company synonymous with the sluggish, seat-licensed world of enterprise CRM, just reported a 200% quarter-over-quarter surge in its AI agent business. I’ve been tracking the AI-crypto convergence from the front lines for a while now, and this isn't just another earnings beat. This is a structural shift in how the enterprise buys software. Chasing the alpha, one block at a time, but this time, the block is a business process, and the alpha is in the orchestration layer.
The market is treating this as a bullish signal for AI adoption, and it is. But I'm digging deeper. We're not just seeing a company sell more software. We are watching the death of the per-seat license and the birth of the per-conversation tax. The sprint never stops, only the pace. And right now, the pace is being dictated by a new pricing model that could either become the industry standard or blow up in Salesforce's face.
Let's cut through the hype and get to the architecture. Agentforce isn't a miracle model. It's an aggregator. It sits on top of the Salesforce CRM stack, not to replace it, but to automate it. The core engine is the Atlas Reasoning Engine, a sophisticated piece of middleware that doesn't train models but routes requests. It taps into a cocktail of external LLMs—OpenAI, Anthropic, Google—and then maps their outputs onto a library of "Atomic Actions." These actions are the real innovation: they are the bridge between raw language and structured CRM commands. Instead of a chat bot that talks, Agentforce is a digital worker that does. It updates records, creates tickets, and schedules meetings.
This "combination-level innovation" is what makes it dangerous. They aren't competing on the SOTA benchmark charts; they're competing on the integration layer. My software engineering background tells me this is where the value is. Model performance is a commodity; workflow reliability is a moat. And that moat is reinforced by their Data Cloud. Agentforce isn't just guessing; it's pulling real-time customer records, order histories, and service tickets to inform its actions. This is a proprietary data flywheel that a generic chatbot like ChatGPT simply cannot access. It's the difference between a generalist intern and a specialist who knows your entire history. From the front lines of the hype cycle, I can tell you this data access layer is the real product.
The commercial pivot is the sharpest part of this knife. Salesforce has moved from a seat-based license to a per-conversation price point, set at $2 per conversation. This is a massive bet. They are explicitly telling the market: we are no longer selling access to software, we are selling successful outcomes. The risk, of course, is that if the AI agent fails, the client doesn't just feel a feature is broken; they watch their bottom line bleed. This "outcome-based pricing" is a double-edged sword. It aligns incentives in theory, but it puts a massive target on Salesforce's back if the "task completion rate" isn't perfect.
When we look at the broader ecosystem, this changes the game. Microsoft's Copilot is a seat-based tool for general productivity. ServiceNow's AI is great for IT workflows. But Salesforce is attacking the revenue-generating core of a company: customer service and sales. They're not just selling a tool; they're selling a "digital labor" to replace the entry-level workforce. The impact on the labor market is something we can't ignore. The "200% growth" figure is an aggregate, and it hides the granularity. The absolute number could be small. If the previous base was minuscule, 200% growth is just a rounding error compared to their $37 billion revenue base. This is the contrarian angle that nobody is talking about: the numbers are flashy, but they haven't changed the fundamentals of the total addressable market yet.
Let's look at the hidden costs. The "per-conversation" model is a beautiful narrative, but it's a nightmare for the client's finance team. With a subscription, you can forecast. With the per-dial tax, you can't. The client might pay $2 for a conversation that requires 10 turns of context. That's $20 for a single interaction. The negative incentive is clear: a poorly designed agent will drive up costs by failing repeatedly, generating more conversations to fix the failures. This is a potential "death by a thousand cuts" scenario for the customer. I think there's a huge hidden risk that the "200% growth" is actually a metric of the user's AI agents failing, and thus generating more billable conversations. The pricing model is a genius revenue generation, but it might be a perverse incentive that undermines the user's trust in the long run.
Also, I'm seeing a gap in the security discussions. We talk about AI alignment, but Salesforce is pushing "business rule alignment." The agent is constrained by your company's predefined workflows. This sacrifices flexibility for safety. In practice, this means you can't use the agent to do anything truly innovative or complex; it's just an automated execution of a pre-approved playbook. The security infrastructure—the Einstein Trust Layer—is designed to handle the sanitization of sensitive data and prevent prompt injection. But this is a black box. We have no independent audits of the security layer. It's the "enterprise trust" narrative, but it's a black-box solution. I'd like to see a third-party audit before I call it a fully secure system.
Let's talk about the competitive landscape. Microsoft Copilot is the elephant in the room. They have the distribution advantage via Office 365. But they are a generalist. Salesforce is a specialist. The developer ecosystem is a huge asset. With over 15 million developers, they can build a massive library of "Atomic Actions" for every possible business process. This creates a deep, defensible ecosystem that is hard for a newcomer to replicate. However, the AI-native startups like Sierra and Intercom's Fin are coming. They are born with the AI-native DNA, and they are building from the ground up without the legacy of the SaaS model. They can be more flexible. The question is whether they can overcome the switching costs and data silos that Salesforce owns.
The 2024 ETF approval taught me that the infrastructure is the base. In the crypto world, the infrastructure is the consensus and the gas. In the enterprise world, the infrastructure is the data. Salesforce's infrastructure is not the compute; it's the data. It's the proprietary data that they own. This is a different kind of moat. And the "200% growth" is a story about the monetization of this data. But the question is: what's the gross margin of this business? The $2 per conversation has to cover the cost of the external LLM APIs, the cloud compute, and the agent's own logic. The cost of a complex conversation might be more than $2, squeezing the margins. It's a high-volume, low-margin play, not the high-margin software play of the past.
The labor market impact is the unspoken elephant. If this works, we will see mass displacement of customer service representatives. This will trigger a massive social backlash. The "AI for the people" narrative is fine, but the "AI for the CFO" is going to be seen as a tool for layoffs. This is the risk that Salesforce might not be prepared for. They are preparing for the technical risks, but not the public relations risk. The story of the 200% growth is a story of a new business model, but it's also a story of an economic shift. We are moving from the "knowledge worker" era to the "prompt worker" era. The next generation of workers will be the ones who know how to build and optimize these agents, not the ones who perform the tasks the agents are replacing.
I want to focus on the "information gain" here. The contrarian insight is that this isn't about AI intelligence at all. It's about pricing the execution of a task. The LLM is a commodity; the execution is the value. The 200% growth is a testament to the fact that enterprises are willing to pay for a completed action, not a conversation. The software is becoming a service provider, not a product. Salesforce is not an AI company; it's a labor provider. And the per-conversation price is the wage of the digital worker. The market hasn't priced this correctly yet. They are pricing the AI, but they should be pricing the displacement of the human labor.
We have to watch the next quarter's earnings call. The number to look for isn't just the "Agentforce bookings," it's the "Customer Success Rate." We need to know the absolute revenue and, more importantly, the customer churn. If the agent is failing to complete tasks, the churn will be high. And the "200% growth" will be a pump, not a trend. The real test is whether the per-conversation model is sustainable. Are they going to offer a hybrid pricing model, a subscription base plus usage? Or will they hold the line on the pure per-dial tax?
Surviving the winter to plant for spring. This is the first real shot of the spring of the digital labor. We are in a sideways market right now, but this is a signal that the enterprise is finally ready to pay for the AI. The question is whether Salesforce can survive the friction of its own innovation. The per-conversation pricing is a revolution, but it's also a threat. They are not just selling software; they are selling a promise. And if they can't deliver on that promise, the 200% growth will be a historical footnote. The sprint never stops, only the pace. I'm watching the pace of the quarterly report, and I'm watching the margin of the per-conversation price. Speed is the only currency that matters, but accuracy is the bank that holds it. The real test is not the revenue growth; it's the unit economics. We're not looking at a company; we're looking at a new financial structure. The question is not if the AI will take the job; it's who will get paid for the output. The answer, for now, is Salesforce. But the question is for how long.
