The $2 Conversation: What Agentforce's Unit Economics Reveal About Crypto's Agent Economy

Salesforce charges roughly two dollars for one AI agent conversation. Not a seat. Not a license. A conversation. The pricing page is public, which is the first thing worth noting, because almost nothing else in this story is. Agentforce bills through Flex Credits — approximately $0.10 per credit, roughly 20 credits for a standard agent interaction. Do the arithmetic and you land near $2.00 per session. That number is the entire ballgame, and it sits in plain sight while an industry argues about whether enterprise agents have finally graduated from demo to revenue. Decoding the signal from the narrative noise requires separating two questions the market keeps fusing: did the technology reach production, and did the business model reach validation? The first has a partial yes. The second has a number nobody will publish. For readers whose frame of reference is on-chain, the enterprise-agent story looks like a foreign country. It is not. The core question inside Agentforce is the same question that has defined crypto's own infrastructure debates for a decade: who captures the value when software executes work autonomously? Salesforce's position is unusual. It does not train frontier models. Agentforce runs on a model-agnostic stack — Anthropic, OpenAI, Google, swapped in and out — wrapped around three proprietary layers: Data Cloud as the factual substrate, the Einstein Trust Layer as the data boundary, and the Atlas Reasoning Engine as the orchestrator. The moat is not intelligence. The moat is integration. One timing note before the analysis. "Dreamforce 2026" as a phrase either describes a forward-looking preview or mislabels the October 2025 edition, whose stated theme was the Agentic Enterprise. Either way the analytical question is identical, and the answer is structurally delayed by three to nine months of revenue recognition lag. That distinction maps directly onto the crypto argument I have been making since the ICO era. In 2017 I ran a three-analyst due diligence sprint across more than fifty whitepapers, and the pattern that killed the majority of those projects was never bad technology. It was the absence of a defensible distribution surface. Nine years later, the enterprise agent market is replaying the same error at enterprise scale, except the participants are trillion-dollar incumbents and the failure mode is quieter. Start with the accounting, because the accounting is the analysis. Salesforce does not disclose Agentforce revenue as a standalone line item. It folds AI results into "Data Cloud & AI" or an "AI & Data Cloud ARR" aggregate. In my work building quarterly narrative risk reports for institutional portfolio managers in 2025, this is the single most reliable tell in software reporting. Companies that have a breakout number publish it. Companies that have a promising number publish it with a growth rate. Companies that have neither publish an aggregate and a customer count. Unearthing the logic within the speculative fog means reading that structure honestly. This is not a claim of fraud. It is a claim about incentive alignment, and incentive structures are where I do my reading. The absence of separate disclosure is itself a disclosure. There is also a definitional gap every reader should hold onto. "Adopt" and "pay at scale" are separated by an order of magnitude. Enterprise software POC-to-production conversion historically runs in the 20% to 40% band; agent products, which require data remediation and workflow redesign before they do anything useful, should be expected to run below that. Press releases count adopters. Invoices count revenue. The distance between those two nouns is where the entire narrative lives. Then run the unit economics. At $2.00 per session, with underlying inference costs plausibly landing between $0.05 and $0.30 depending on model selection and context length, plus orchestration, storage and the compliance overhead of the Trust Layer, gross margin on a mature session is theoretically comfortable — call it 60% to 85%. That is a good business. It is not a good business at low volume. Enterprise agent sessions consume enormous context: retrieval over knowledge bases, permission inheritance checks, multi-step tool calls, audit logging. The token load per interaction is not chatbot-grade. It is an order of magnitude heavier. A customer running two thousand sessions a month is not amortizing the fixed cost of their own implementation. A customer running two million is printing margin. The distribution between those two populations is the number that matters, and it has never been published. The pivot point where genre defines value is right here. Traditional SaaS monetizes seats: a predictable, contractually locked, expansion-friendly unit that investors learned to price at premium multiples because of its retention characteristics. Consumption pricing monetizes events: volatile, unforecastable, and structurally deflationary as models get cheaper. Salesforce is migrating its core revenue model from the first genre to the second while simultaneously selling the AI that accelerates the migration. There is a self-cannibalization problem buried inside that. If an agent replaces three human service seats, the customer buys fewer seats. Salesforce's seat revenue is its legacy core. Which means the entity best positioned to sell agentic replacement is also the entity with the strongest incentive to slow that replacement down. This conflict rarely surfaces in bullish coverage, and it is the most important structural fact in the sector. Now bring in the part crypto has been waiting to hear. Consumption-based billing at machine granularity — $0.10 units, thousands of transactions per hour, no human in the loop — is precisely the use case public blockchain settlement was designed for and has never been able to win. For eight years the pitch has been that stablecoins and high-throughput chains will become the payment rail for machine-to-machine commerce. The enterprise demand is arriving on schedule. It is just not arriving on the rails built for it. Look at what actually clears. Agent payments at enterprise scale settle through card networks, through bank wires on monthly invoices, through committed-use contracts with true-up clauses. The nascent machine-payment standards gaining traction — payment-intent protocols, agent-checkout frameworks, tokenized card credentials — are drafted by the incumbents, for the incumbents. Stablecoins appear in this stack as a settlement efficiency, not an ideological commitment. A treasury team choosing USDC for cross-border agent payouts is not adopting a chain. It is choosing a cheaper wire. The protocols that will define machine commerce are being written right now, and crypto-native entrants are not holding the pen. Agent registries with signed identity documents, spend authorizations bound to user consent, dispute paths that terminate in a card network's chargeback rules — the working implementations ship from payment processors and model labs. On-chain equivalents exist and are technically elegant. Elegance has never won a standards fight. I watched the same dynamic during the 2021 NFT transition from profile pictures to utility assets: the losers were not the projects with worse tech. They were the ones that arrived after the distribution was claimed. The moat Salesforce actually possesses — data gravity plus AppExchange distribution — has nothing to do with settlement layer. Agentforce does not need a public chain to bill $2 per conversation because it already has a billing relationship with the customer. The unglamorous truth of the agent economy is that the hard part was never the payment rail. The hard part is knowing whether the agent has permission to do the thing it is about to do. That is the trust layer question, and it is where the crypto-native framing has been most overconfident. On-chain attestation, verifiable compute, decentralized identity — these address real problems, and enterprises genuinely have them. But the enterprise answer to "how do I know the agent is authorized" is currently a permission-inheritance model inside the CRM, audited by the vendor, backed by a contract that places liability on the customer. It works. It is ugly. It ships. The BPO displacement story follows the same logic. Business process outsourcing is a several-hundred-billion-dollar market, and its core deliverable — tier-one customer interaction — is the most directly substitutable task in the economy. Headline risk is real. But replacement is gated by something the narrative consistently ignores: regulated industries need human-in-the-loop review, which caps substitution rates. Finance, healthcare and legal will run far below marketing and support benchmarks, because the cost of an agent acting wrongly is not a bad sentence. It is a regulatory event. And the demand that does convert will not accrue to the agent platform first. It accrues to whoever cleans the data. In my DeFi Summer mapping work, the least visible layer was always the most valuable — the liquidity plumbing beneath the yield narrative. Enterprise agents have the same shape. Data governance, ETL, master data management: boring, necessary, high-margin prerequisites. They are also where implementation cost overruns occur, which is the mechanism by which a "high signings" quarter becomes a "low activation" year. The competitor set reinforces this. Microsoft owns distribution through Office and Azure. ServiceNow owns process depth. Google owns model capability and Workspace reach. Salesforce owns the customer record. None of them owns the agent standard, and that standard war — the protocol by which agents discover, authenticate and transact with each other — will resolve the way the Layer 2 standard war resolved: not by technical superiority, but by who convinces more participants to deploy first. Here is the position that will cost me some readers. The agent economy's total addressable market, as measured by the enterprises actually writing cheques, will be settled largely outside public blockchain infrastructure — permanently. Not because chains are slow, and not because regulation blocks them. Because the buyer does not have the problem chains solve. An enterprise procurement team running agentic workflows already has a billing relationship, a legal counterparty, a dispute-resolution path and an auditor. Adding a public settlement layer introduces counterparty risk, key management, jurisdictional ambiguity and a new vendor to onboard. The proposition — cheaper, faster, programmable settlement — is real but marginal against what the enterprise already has. I watched this exact dynamic stall real-world-asset tokenization for three years. Institutions did not need the chain. They needed a ledger, and they already had one. The crypto-native opportunity is not the settlement rail for enterprise agents. It is the long tail: agent-to-agent micro-transactions between parties with no pre-existing billing relationship, no credit terms, no invoices. That market is real. It is also, today, a rounding error next to enterprise volume — and it will be underwritten by a story about the enterprise market that the enterprise market has not agreed to. Building frameworks for the next narrative cycle means watching three signals, not one headline. Whether Salesforce breaks out Agentforce revenue as a separate line — if it stays aggregated for another four quarters, read that as narrative cooling, not conservatism. Whether net revenue retention on AI products diverges from the CRM base. And whether seat revenue visibly decelerates while AI consumption accelerates, which is the cannibalization signal and the one that rerates the entire software complex. The agent economy is being built. It is being built on billing systems that already exist, by vendors that already own the customer. The question crypto should be asking is not whether agents will need a settlement layer. It is whether the layer they end up using will have a token.