The $2 Conversation: Salesforce's Agentforce and the DeFi-ification of Enterprise SaaS

The $2 conversation is the most important number in enterprise software right now. And almost nobody is talking about it.

Salesforce dropped its Q2 earnings, and the market narrative is a familiar one: AI is the centerpiece, Agentforce is the rocket, growth is coming. The headlines write themselves. The press releases polish the story. But strip away the marketing layer and you find a single pricing signal that changes the entire unit economics of the SaaS industry. It is not a subscription. It is not a license. It is a per-conversation fee of roughly two dollars.

That number is a structural break from forty years of software economics. And it deserves a forensic teardown.

The Context: From System of Record to System of Action

Let me be precise about what Salesforce is actually building here. Agentforce is not a chatbot. It is not a copilot that suggests replies. It is an autonomous execution layer that sits on top of the CRM stack — Sales Cloud, Service Cloud, Marketing Cloud — and actually completes tasks. It resolves tickets. It follows up on leads. It handles end-to-end customer service requests without human intervention.

The technical architecture is unremarkable. Agentforce is a composable integration — a combination of large language model APIs (OpenAI, Anthropic, Google, depending on the task), the Data Cloud for context, and the Flow workflow engine for orchestration. There is no proprietary foundation model. There is no novel AI breakthrough. The moat, if there is one, is not in the model. It is in the data layer and the workflow integration.

Salesforce holds the largest CRM dataset on the planet. Over 150,000 enterprise customers have spent a decade feeding their sales pipelines, service histories, and customer interactions into this system. That is the fuel. And Agentforce is the engine that finally burns it.

But here is where the story gets interesting. The shift from subscription to usage-based pricing is not just a commercial decision. It is an admission that the old model — charging per seat for software that people may or may not use — is broken. Math has no mercy. If the software does the work, why would the customer pay for the worker?

The Core: Unit Economics of an Autonomous Agent

Let me walk through the numbers, because this is where the analysis gets real.

Agentforce pricing is set at approximately $2 per conversation. A conversation is defined as a completed task — a resolved ticket, a qualified lead, a completed follow-up. The question that matters is: what does it cost Salesforce to deliver that conversation?

I have built enough cost models to know the variables. A typical enterprise customer service interaction consumes roughly 5,000 to 10,000 tokens (input plus output). At current API pricing for GPT-4-class models — roughly $10 to $30 per million tokens — the raw inference cost per conversation lands between $0.05 and $0.30. That leaves a gross margin of 85% to 97%.

On the surface, those are beautiful numbers. But the surface is where naive models live.

There are hidden costs. The Data Cloud retrieval layer adds latency and compute. The workflow orchestration — Flow triggers, API calls to legacy systems, database lookups — consumes infrastructure. The enterprise-grade security layer, the audit trails, the compliance checkpoints... all of it eats into that margin. And then there is the cost of the humans who have to supervise the agents when they inevitably go off the rails.

I am reminded of my 2020 work modeling DeFi yield curves. The headline APYs looked incredible — 20%, 50%, 100% — until you decomposed the token emissions and realized the yields were subsidized by inflation. The underlying protocol was bleeding value with every block. The same principle applies here. The question is not whether the gross margin is 90% at pilot scale. The question is what happens at scale, with real traffic, with edge cases, with adversarial inputs, with the long tail of customer behavior that no training set fully captures.

The second structural issue is the revenue volatility. Subscription revenue is predictable. You sign a contract, you recognize revenue ratably, you can forecast with confidence. Usage-based pricing is inherently spiky. A customer's conversation volume depends on their business cycle, their seasonality, their own customer demand. This introduces variance into the financial model. And variance is the enemy of a 50x earnings multiple.

This is the core tension. Salesforce is trading the stability of the old model for the upside of the new one. That is a rational bet. But it is a bet on execution, not on narrative.

The Contrarian Angle: What the Bulls Get Right

The bear case is easy. I have written enough of those. But let me steelman the bull case, because there is a real argument here that most skeptics miss.

The bull case is not about the technology. It is about the distribution.

Salesforce has 150,000 enterprise customers. They have a trusted relationship. They have an integration footprint that took years to build. When a company has its sales pipeline, its customer service history, its marketing automation all running on Salesforce, switching costs are enormous. And when Salesforce says "add Agentforce to your existing stack," the procurement conversation is dramatically easier than adopting a point solution from an AI-native startup.

The second point is the data flywheel. Every conversation Agentforce handles generates data that improves the model's context. The system gets better at understanding a specific customer's business, their tone, their policies, their exceptions. This is a compounding advantage that pure-play AI companies cannot replicate — they have the models, but they lack the enterprise context.

I saw this dynamic play out in DeFi in 2021. The protocols that won were not the ones with the most innovative code. They were the ones with the deepest liquidity and the most entrenched user relationships. The technology was a commodity. The distribution was the moat.

That said, the comparison cuts both ways. In DeFi, the projects with the deepest distribution often had the worst tokenomics. The incentive structures were misaligned. And when the market turned, the users left and the liquidity evaporated. The same risk exists here. If Agentforce's per-conversation pricing does not deliver measurable ROI for customers, the adoption curve will stall, and the narrative will collapse.

The third point in the bull case is the competitive positioning. Microsoft Copilot is the obvious threat. But Microsoft's approach is fundamentally different. Copilot is embedded in the productivity suite — it helps you write emails, summarize documents, automate Excel. It is a tool for knowledge workers. Agentforce is a tool for operations. It replaces the worker, not just assists them. These are different markets with different buying dynamics.

And Salesforce has a data advantage. Dynamics 365 is a distant second in CRM market share. The data that matters — the actual customer interaction history — lives in Salesforce. This is a structural advantage that will be difficult for Microsoft to overcome, regardless of their Azure cost advantages.

The Takeaway: Verify the Stack

Here is my honest assessment. The Agentforce story is the most important commercial experiment in enterprise AI right now. Not because the technology is revolutionary — it is not. But because it is the first serious test of whether usage-based pricing for autonomous agents can work at scale.

I trust, verify the stack. And the stack here has a critical vulnerability.

The per-conversation pricing model creates an inherent conflict of interest. Salesforce has an incentive to maximize the number of conversations, not the quality of outcomes. The customer has an incentive to minimize conversations while maximizing value. This misalignment will create friction, and the question is whether the pricing model evolves to become outcome-based rather than activity-based.

If Agentforce succeeds, it will redefine the SaaS industry. The subscription model will be under pressure across every category. If it fails, it will be a cautionary tale about the gap between AI hype and enterprise reality.

I am watching three specific signals. First, the gross margin trend — if Salesforce discloses AI-related cost structures, the unit economics will tell us whether the $2 price is sustainable. Second, the Net Revenue Retention rate for Agentforce customers — are they expanding usage or churning? Third, the number of conversations per customer — is this a high-frequency engagement or a novelty that fades after the pilot?

The enterprise AI land grab is underway. Salesforce has made its move. The next four to six quarters will determine whether this is a paradigm shift or a very expensive experiment. High yield, high graveyard. The same law applies to enterprise software as to DeFi. The question is not whether the agents can think. It is whether the economics can survive contact with reality.

Rug pulls are just bad code. And the code here — the pricing model, the incentive structure, the unit economics — is still being written. I will be watching the next earnings call with a calculator and a healthy dose of skepticism.

Math has no mercy. And neither do the markets when the narrative meets the numbers.