Hook: The Metric That Demands an Audit
Salesforce reported that its AI agent business, Agentforce, grew over 200% year-over-year. The market nodded approvingly. The stock ticked up. The narrative solidified: enterprise AI has found its killer app.
I've spent 25 years reading earnings call transcripts the way forensic accountants read ledgers. And this particular number—200%—triggers every alarm I have. Because in my experience, when a company leads with a percentage instead of an absolute figure, there's usually a reason. The ledger never lies, only the interpreter does.
Let me be precise about what we know. Salesforce's Agentforce is an AI agent platform integrated into its CRM suite. It handles customer service inquiries, sales lead qualification, and marketing tasks. The pricing model is $2 per conversation. The growth rate is "over 200%." That's the entire public data set.
What we don't know: the revenue base from which that 200% is calculated. The customer count. The retention rate. The gross margin. The cost per conversation. The churn. The absolute dollar contribution to Salesforce's roughly $37 billion in annual revenue.
Without those numbers, "200% growth" is a whisper, not a signal. And in the absence of noise, the signal screams.
Context: What Agentforce Actually Is
Agentforce is not a foundation model company. It doesn't train large language models. It doesn't claim to. What it does is orchestrate them.
The architecture is built on the Atlas Reasoning Engine, which routes queries to external models from OpenAI, Anthropic, and Google. The outputs are mapped to Salesforce's CRM objects through what the company calls "Atomic Actions"—pre-built functions that translate model responses into business operations like updating a customer record, creating a service ticket, or triggering an email.
This is integration innovation, not model innovation. The technical moat is not in the intelligence—it's in the plumbing. The data access layer, powered by Salesforce Data Cloud, gives Agentforce real-time access to structured business data: customer histories, order records, service interactions. That's commercially more valuable than the public data general models train on.
The Einstein Trust Layer sits on top, handling data masking, prompt injection protection, and audit trails. For enterprise customers, this is the difference between "interesting demo" and "production deployment."
The technology is real. The engineering is competent. The question is whether the business model holds up under stress.
Core: The $2 Conversation Economics
Let me walk through the unit economics, because this is where the story gets uncomfortable.
Agentforce charges $2 per conversation. That conversation requires a call to an external LLM API. The cost of that call depends on the model, the token count, and the complexity of the task. A simple order status inquiry might cost $0.05 in inference. A complex multi-step troubleshooting session could cost $0.50 or more.
The gross margin on each conversation is therefore variable. At scale, with negotiated bulk pricing from model providers, Salesforce might achieve 70-80% gross margins on the inference cost alone. But that's before accounting for the infrastructure to run the orchestration layer, the Data Cloud integration, the Trust Layer security, and the customer success teams.
Here's the structural problem: the pricing model transfers risk from the customer to Salesforce. Under traditional SaaS, customers pay per seat regardless of usage. Under Agentforce's model, customers pay per conversation. If the AI agent fails to resolve an issue, the customer doesn't pay for a successful outcome—they pay for the failed attempts too. Every failed conversation is revenue for Salesforce but zero value for the customer.
This creates a perverse incentive structure. The more the AI agent struggles, the more conversations it generates, the more revenue Salesforce books. But the customer sees rising costs with no corresponding improvement in outcomes. That's not a sustainable value proposition. That's a churn machine.
Based on my experience auditing the MakerDAO stability fee model in 2020, I can tell you exactly how this plays out. When a pricing mechanism doesn't align with actual value delivery, the market corrects it—brutally and without warning.
The 200% growth number, viewed through this lens, becomes less impressive. If the base was small—say $50 million in annualized revenue—200% growth brings it to $150 million. That's 0.4% of Salesforce's total revenue. It's a rounding error dressed up as a strategic pivot.
The Data Moat: Real But Narrow
Salesforce's genuine advantage is the data layer. The Data Cloud integration means Agentforce can access customer records, order histories, and service tickets in real time. This is the moat. A competitor building an AI agent from scratch would need to replicate years of accumulated CRM data and the integrations that connect it.
But here's the contrarian angle: data moats are only as strong as the workflows they support. If the Atomic Actions library doesn't cover a customer's specific long-tail business processes, the customer must build custom integrations. That's expensive, time-consuming, and requires specialized skills that most enterprise IT teams don't have.
The hidden cost of Agentforce adoption is the implementation complexity. The 200% growth likely comes from existing CRM customers who are already invested in the Salesforce ecosystem. They're not choosing Agentforce because it's the best AI agent—they're choosing it because it's the path of least resistance. That's penetration, not conquest.
Competition: The Wolves at the Gate
Microsoft Copilot is priced per seat, which feels familiar to enterprise buyers. ServiceNow's AI agents are priced per workflow, which aligns with IT service management processes. Both are attacking from different angles.
Microsoft has the Office 365 distribution advantage. Every enterprise already pays for Microsoft. Adding Copilot is a line item, not a new vendor. ServiceNow has the IT service management incumbency. Their AI agents are purpose-built for the workflows that IT departments already run.
Salesforce's defense is the depth of CRM integration. But that's a narrowing moat. The AI-native startups—Sierra, Intercom's Fin, and others—are building from scratch without legacy constraints. They don't have to integrate with decades of accumulated technical debt. They can design for AI from day one.
The competitive pressure is real, and it's intensifying. Salesforce's 200% growth, if it's real, is happening in a market where every major platform player is spending aggressively. The question isn't whether Salesforce can grow—it's whether it can grow profitably and sustainably against better-funded, more focused competitors.
The Regulatory Shadow
The EU AI Act will likely classify customer service AI agents as high-risk systems. That means transparency requirements, human oversight obligations, and documentation standards. Salesforce has said it will comply, but compliance is expensive.
The deeper issue is accountability. When an AI agent makes a wrong decision—an improper refund, a misleading promise, a data breach—who's responsible? Salesforce? The customer? The model provider? The legal framework is undefined. In my analysis of the Terra/Luna collapse, I saw what happens when responsibility is diffuse: everyone blames everyone else, and the users absorb the loss.
For enterprise customers, this ambiguity is a deal-breaker in regulated industries. Financial services, healthcare, and government agencies will move slowly. The 200% growth is likely concentrated in less-regulated sectors—retail, technology, professional services. The total addressable market is smaller than the narrative suggests.
The Infrastructure Reality
Salesforce doesn't train foundation models. Its compute needs are inference-heavy, not training-heavy. This is a strategic advantage—it avoids the massive capital expenditure of model training. But it creates a dependency on external model providers.
If OpenAI or Anthropic raises prices, Salesforce's margins compress. If a model's capabilities plateau, Agentforce's performance stagnates. If a model provider has an outage, Agentforce goes down with it.
The negotiation leverage is real—Salesforce is a large customer—but the dependency is structural. This is not a moat; it's a lease.
Contrarian: The 200% Growth Is a Distraction
Here's what the market is missing. The 200% growth number is designed to signal AI leadership. It's a narrative tool, not a financial metric. The real questions are:
What's the absolute revenue? What's the gross margin? What's the customer retention rate? What's the cost per conversation? What's the implementation time? What's the churn rate?
Without those numbers, the 200% is noise. And in my experience, when a company leads with a percentage instead of an absolute, the absolute is usually small.
The more important signal is the pricing model change. Moving from per-seat to per-conversation is a fundamental shift in how SaaS companies think about value. If Salesforce can make this work, it will reshape the industry. If it fails, it will be a cautionary tale about the dangers of outcome-based pricing without guaranteed outcomes.
The Data Flywheel: Real But Slow
The genuine opportunity is the data flywheel. Every Agentforce interaction generates data about customer behavior, common issues, and effective resolutions. This data can improve the Atomic Actions library, making the platform smarter over time. Competitors can't replicate this without similar deployment scale.
But the flywheel is slow. It takes months of production data to generate meaningful improvements. And the improvements are incremental, not transformative. The moat is real, but it's a slow-growing one.
Takeaway: What to Watch
The next quarter's earnings call will be the first test. I'll be looking for three numbers: Agentforce's absolute revenue, its gross margin, and its customer count. If Salesforce reports these numbers transparently, the 200% growth is credible. If it continues to lead with percentages, the growth is likely a mirage.
The second signal is pricing. If Salesforce introduces a hybrid model—base subscription plus usage-based fees—it's acknowledging the risk in pure per-conversation pricing. That would be a smart move, but it would also be an admission that the current model has problems.
The third signal is customer churn. If enterprise customers are renewing at high rates, the model works. If not, the 200% growth is a one-time spike, not a sustainable trend.
Whales don't announce their positions. They accumulate quietly and move when the time is right. The same logic applies to enterprise AI adoption. The real signal isn't the growth percentage—it's the retention rate, the margin, and the absolute revenue.
Correlation is a whisper; causation is the shout. The 200% growth is a correlation. The causation is in the unit economics, the competitive dynamics, and the regulatory environment. That's where the truth lives.
The ledger never lies, only the interpreter does. And right now, the market is interpreting a percentage without examining the ledger behind it.
In the absence of noise, the signal screams. The signal here is that Salesforce has built a competent AI orchestration layer with a real data moat. But the business model is unproven, the competition is intensifying, and the regulatory environment is uncertain. The 200% growth is a starting point, not a conclusion.
The next twelve months will determine whether Agentforce is a genuine transformation or a well-marketed experiment. I'll be watching the numbers, not the narrative. Because in the end, the data always tells the truth. It's just a matter of whether anyone is listening.
The Final Word
Salesforce's Agentforce is a real product with real engineering and a real data advantage. But the 200% growth figure is a narrative device, not a financial disclosure. The per-conversation pricing model is a bold experiment that could reshape SaaS economics—or collapse under the weight of its own incentives.
The market should demand more transparency. The absolute numbers. The margins. The retention rates. The churn. Without those, the 200% is just a number designed to move a stock price.
I've seen this pattern before. In 2021, I tracked a CryptoPunks whale who was inflating floor prices through wash trading. The volume looked impressive. The reality was self-dealing. The same logic applies here: growth that can't be verified is growth that should be questioned.
The next earnings call will tell us more than this article ever could. I'll be there, spreadsheet open, checking the numbers against the narrative. Because in the end, the data doesn't lie. It just waits for someone to read it carefully.
And I always read carefully.
