Roughly 3,400 words arrived on my desk last week labeled as blockchain and Web3 reporting. I ran the standard intake check: count of transaction hashes, count of contract addresses, count of block numbers, count of RPC calls. All four returned zero. The subject was enterprise software. Salesforce. A product family called Agentforce, deployed under a marketing umbrella the source rendered as "AIforce."
The document contained one number that belonged in my discipline: $2 per conversation.
I have spent four years reconciling consumption meters. In 2026 I mapped IP-to-wallet correlations across a cluster of autonomous trading bots and published the pattern-recognition logic that exposed a $10 million wash-trading scheme. The unit of analysis in that investigation was not the bot. It was the meter that priced the bot's work. The same question applies here, and it is the only question worth asking about a press release that wandered into the wrong newsroom: who set the meter, what does it measure, and can an outside party verify it.
Tracing the source produced three findings, two of which are commercially useful and one of which is a warning about the source itself.
The product is real, and I want to be precise about the line between verified and unverified before I proceed.
Salesforce operates Agentforce, an enterprise agent platform. It meters usage under published rates that include a $2 per conversation track and a credit denomination called Flex Credits. It has a commercial relationship with Anthropic, and Claude supplies the underlying model. The company has adopted Model Context Protocol, the open specification Anthropic published in November 2024, and it exposes Salesforce objects through a headless toolkit that reaches external front ends including Slack, Lightning, and the Claude application. Those elements reconcile against public information. I treat them as confirmed.
Four elements do not reconcile. The source dates the announcement to "Dreamforce 2026." Dreamforce runs in the autumn; a 2026 event described in the past tense is a calendar contradiction. The brand names "AIforce," "Claudeforce," and "Slackforce" do not appear in any Salesforce documentation I can retrieve. The publicly documented agent brand is Agentforce, extended as Agentforce 360.
My working hypothesis, stated plainly because unstated hypotheses are how bad analysis propagates: the source is a rewrite or translation of Dreamforce 2025 material, with substituted brand names and a corrupted date field. The commercial data survived the rewrite. The identifiers did not.
That distinction governs everything below. Pricing survives; provenance does not. I will analyze the first and flag the second.
Follow the outflows and the architecture becomes legible. Agentforce is not a model. Salesforce does not train a foundation model. It rents one and sells orchestration. The technical center of gravity is MCP plus headless API exposure, which places the work in the engineering and integration tier, not the model tier. MCP is not a firewall and not a sandbox; it is a standardized wrapper for tool invocation and context injection. Its function is to let an external model read and write against enterprise records under an inherited permission model.
Three claims in that architecture require verification and receive none.
First, Zero Data Retention. The source reports that customer data is neither stored nor used to train external models. This is the correct commitment. It is also a claim about the interior of an inference pipeline, which no external party can inspect from outside. Prompt assembly, retrieval buffers, and transient caches are precisely the layers where retention commitments historically degrade. I have audited tokenized real estate under MiCA and found two of three projects failing proof-of-reserve standards because custodial relationships were opaque. The failure mode is identical here: a strong commitment, a documented intent, and no independent attestation of the code path.
Second, governance continuity. The source asserts that existing permission models extend automatically across Claude, Slack, and Lightning without relearning. Cross-system permission equivalence is the hardest problem in enterprise agent deployment. When an agent acts on behalf of a user and traverses object boundaries, the effective permission set is a union of contexts. Unions leak. No technical validation is provided.
Third, agent reliability. The source cites 100,000 user activations in 35 days. That is an activation metric, not a paid conversion metric and not an effective utilization metric. It measures account provisioning. It does not measure whether a multi-step workflow completed end to end without human intervention. This is the same reporting failure I documented in 2022, when the Terra collapse was narrated as sentiment while the actual causal chain โ 14,000 wallets, a structural peg failure โ was visible only in the flow data. Activations are the sentiment metric here. Task completion rate is the flow data. It is absent.
Now the meter, which is where this material becomes genuinely instructive.
Salesforce meters under four tracks. A low tier at $5 per user per month. A high tier starting at $550 per user per month. A consumption track at $2 per conversation. And a credit denomination at $500 per 100,000 Flex Credits.
The crossover is the whole story. At $550 per seat, a customer reaches parity with consumption billing at 275 conversations per month. Below 275, seats are cheaper. Above 275, the consumption track dominates. That threshold is not a coincidence and not a rounding artifact. It is a value anchor placed deliberately at the point where an organization must decide whether its agents are occasional assistants or primary labor.
This is a metering problem I recognize. On-chain, gas is the meter, and every protocol that has tried to abstract gas behind a credit system has paid for it in transparency. Flex Credits perform exactly that abstraction. They insulate the customer from token counts and conversation counts alike, which is convenient in the short term and advantageous for the vendor in the long term. A meter denominated in a proprietary credit can be repriced, rebundled, or retired without the customer's invoice changing shape. The source's own total cost estimate for a 30-seat team lands between $200,000 and $450,000 in year one, inclusive of Service Cloud Enterprise Edition, implementation, and credits. That is $6,667 to $15,000 per seat annually. The meter is not a discount channel. It is a new revenue tier.
The implementation line inside that estimate deserves separate treatment, because it is load-bearing. A year-one range of $200,000 to $450,000 for thirty seats explicitly bundles integration services. Agent deployment is not self-service. It requires permission modeling, object mapping, workflow definition, and failure-path design. That work lands on systems integrators, and it recreates the services economy that surrounded enterprise resource planning two decades ago. The direct beneficiaries of an agent platform launch are frequently not the platform vendor's shareholders. They are the consultants who make the platform functional.
Here is the part that transfers directly to my discipline. A meter that prices autonomous work will be gamed by whoever controls the agent. I proved this in 2026 with trading bots. The bots did not break the exchange; they exploited a metering model that counted activity rather than value. Consumption pricing on enterprise agents creates the same incentive surface. There is no reported cost ceiling in the source material. An agent in a retry loop is an unbounded invoice. The technical risk of agent failure converts into a financial risk of billing failure, and the second is easier to trigger than the first.
The interface question is the strategic one, and it is the one the source never raises. Salesforce exposes its data through MCP to Claude, to Slack, to Lightning. That is openness, and it is also a relocation of the user interface. If the user's mental model of the product becomes "Claude with my Salesforce data," Salesforce has become a context supplier to a front end it does not control. The brand names in the source โ Claudeforce, Slackforce โ read as an attempt to claim co-ownership of that interface. Whether the claim holds is a distribution question, not a naming question, and the answer depends on where usage actually concentrates across those three surfaces. The source provides no distribution data.
The obvious inference from four-track pricing is that per-seat software is dying and consumption is its successor. I want to separate that inference from the evidence.
Correlation is not causation, and a vendor offering multiple meters is not proof that any one of them is winning. A four-track structure is best read as hedging, not as a declaration. Salesforce is retaining seat-based revenue for budget-constrained accounts while building a consumption track for automation-heavy accounts. Both can coexist for years. The strategic admission โ that seats no longer measure value accurately once agents perform the work โ is real, and it is significant. The strategic outcome โ that seats decline and consumption compensates โ is unproven.
The measurement that would settle it does not exist in the source. Net revenue retention split by billing track. Paid conversion against the 100,000 activations. Gross margin per conversation, net of Claude inference cost. The source discloses a price and withholds a cost. In seat-based software, marginal cost approaches zero and gross margin is trivial. In agent software, every interaction consumes rented inference. A $2 conversation that costs $1.40 to serve is a structurally worse business than a $2 seat that costs almost nothing to serve, and the source provides no way to distinguish the two cases.
There is a second-order effect worth flagging for anyone holding infrastructure exposure. This pricing model creates demand for cost observability, agent audit trails, and permission verification as separately purchased products. Seat pricing generated no such market. Consumption agent pricing generates it immediately, and I have a professional interest in that market because I build inside it.
The source itself deserves one line in the findings. A document labeled blockchain and Web3, containing no blockchain content, is a provenance failure rather than a reporting failure. Aggregators relabel content to satisfy feed taxonomies. The labels then propagate into search indexes and citation chains. I flag it because source integrity has been a precondition of my work since 2021, when I submitted a 50-page discrepancy report to protocol repositories rather than to social feeds, and the engineers who read it cared about exactly one property: whether they could re-derive my numbers from the referenced blocks. A reader of the Agentforce source cannot re-derive the activations, the retention behavior, or the margin. The ledger doesn't flatter anyone, and an article that cannot be reconstructed is not evidence.
Watch three signals. Whether Salesforce reports consumption revenue under a separately disclosed line, because a vendor that considers a meter strategic names it in the financials. Whether any third party attests to Zero Data Retention at the pipeline level, because an unaudited commitment is a policy rather than a control. And whether Microsoft or ServiceNow publishes a consumption track of its own within four quarters, because a single vendor's meter can be repositioned while an industry's meter cannot.
The 275-conversation crossover point is the number I will be tracking. Audit complete, for now.
