Salesforce's Agentforce: The Per-Conversation Tax on Enterprise AI

The market is celebrating a 200% growth number. I am more interested in the unit economics that make it possible. Salesforce's Agentforce has crossed the threshold from demo-ware to production, but the pricing model that fuels its expansion contains a structural flaw that most enterprise software analysts are missing. This is not a story about AI capability. It is a story about liquidity, incentive alignment, and the hidden costs of outcome-based pricing in a market that has never been forced to price for certainty. When I audited ICO whitepapers in 2017, I learned to look for the mechanism that breaks first. The same principle applies here. Agentforce is not a model company. It is an orchestration layer that routes queries to OpenAI, Anthropic, and Google models, then maps the outputs onto CRM objects through what Salesforce calls Atomic Actions. The technical architecture is sound. The business model is the vulnerability. Salesforce has adopted a per-conversation pricing model at $2 per dialogue. This is a fundamental departure from the per-seat licensing that has defined SaaS for two decades. The shift from charging for access to charging for outcomes is philosophically correct. It aligns vendor revenue with customer value. But it also transfers all execution risk from the buyer to the seller. If the AI agent fails to resolve a customer issue in three attempts, the customer pays for three conversations and receives zero value. The cost is not absorbed by the enterprise. It is absorbed by Salesforce in the form of churn. This is the same maturity mismatch I identified in Compound Finance in 2020. The protocol looked healthy because TVL was growing. The interest rate curves were unsustainable below 150% collateralization. The market was pricing for continuation, not for stress. Agentforce's 200% growth is a TVL number. It tells us nothing about the retention rate, the average conversation completion rate, or the gross margin after inference costs. Without those data points, the growth figure is a narrative, not a signal. The data moat is real. Salesforce's Data Cloud provides access to structured business data that no general-purpose model can replicate. Customer records, order histories, and service tickets are proprietary assets. This creates a switching cost that is genuinely difficult to overcome. But the moat is in the data layer, not in the AI layer. Any competitor with access to similar CRM data can build a comparable orchestration stack. Microsoft Copilot has the Office 365 distribution advantage. ServiceNow has the IT service management workflow advantage. The differentiation is thinning. My concern is the inference cost structure. Agentforce does not train models. It rents them. Every conversation incurs a variable cost that scales linearly with usage. At $2 per conversation, the gross margin depends entirely on the wholesale price Salesforce negotiates with model providers. If the blended inference cost is $0.50 per conversation, the margin is healthy. If it is $1.50, the business is barely break-even. The market does not know this number. Salesforce has not disclosed it. This opacity is the enemy of alpha. Volatility is the tax on unproven consensus. The consensus here is that Agentforce represents a new category of digital labor. The unproven assumption is that enterprises will accept a pricing model that penalizes them for AI failure. In traditional SaaS, the vendor bears the cost of software bugs. In per-conversation pricing, the customer bears the cost of AI hallucination. This is a fundamental inversion of risk allocation. It will work in a bull market for AI adoption, where enterprises are experimenting with budgets. It will break in a bear market, when CFOs scrutinize every line item and demand ROI guarantees. The contrarian angle is that Salesforce's growth is a base effect. If the prior year's revenue from Agentforce was negligible, a 200% increase still represents a small absolute number. The company's total revenue is approximately $37 billion. Agentforce's contribution is likely a rounding error. The market is pricing Salesforce as an AI leader based on narrative momentum, not on financial materiality. This is the same pattern I observed with Terra's 20% APY. The mechanism was unsustainable, but the market extrapolated the growth curve without questioning the base. There is also a labor market dimension that the market is ignoring. Agentforce is designed to replace customer service representatives, telemarketers, and junior marketing specialists. This is not a hypothetical. The technology is production-ready. The social and regulatory backlash will be significant. The EU AI Act will likely classify customer service AI as high-risk, requiring transparency and human oversight. Salesforce has announced compliance, but the cost of compliance will erode the unit economics. The per-conversation price will need to rise, or the margin will compress. The infrastructure dependency is another hidden variable. Salesforce relies on AWS, Azure, and GCP for GPU capacity. It has no chip design capability and no training infrastructure. This is a rational strategy, but it creates a dependency on the pricing power of cloud providers and NVIDIA. If GPU prices rise, the inference cost increases, and the $2 per-conversation price becomes untenable. Salesforce has negotiating leverage as a large customer, but it does not have pricing power over the underlying commodity. My framework for evaluating this is the same one I used for the 2024 ETF arbitrage. The basis trade worked because the spread was predictable and the risk was contained. Agentforce's per-conversation model is the opposite. The spread between the price and the cost of delivery is unknown. The risk is not contained. It is a function of model performance, which is outside Salesforce's control. The company is essentially writing a put option on the reliability of third-party AI models. If the models improve, the margin expands. If they stagnate, the margin collapses. The takeaway is not that Agentforce will fail. It is that the market is pricing for a certainty that does not exist. The 200% growth number is a headline. The retention rate, the gross margin, and the absolute revenue contribution are the metrics that matter. Until Salesforce discloses these numbers, the AI premium in its valuation is a bet on narrative, not on fundamentals. I am watching the next earnings call for the unit economics. The chart tells the truth the tweet hides. The per-conversation model is a tax on unproven consensus, and the market is paying it willingly.