A four-tier price list published in October by the largest customer-relationship-software vendor in the world marked the first time a top-ten enterprise software company priced its flagship product by the unit of work performed instead of the number of humans granted access.

Five dollars per user per month at the floor. Five hundred and fifty dollars per user per month at the ceiling. In between sits a metered track: two dollars per conversation, or five hundred dollars for one hundred thousand credits denominated in the vendor's own unit of account.

The announcement did not publish the arithmetic underneath those numbers, so I ran it. A multi-step agent conversation — meeting preparation, deal-health review, pipeline update — plausibly consumes 20,000 to 40,000 tokens across a reasoning chain that includes tool calls, retrieval, and a verification pass. At frontier API rates near $3 per million input tokens and $15 per million output, a 30,000-token chain with a four-to-one input-to-output split costs roughly $0.12 of input and $0.27 of output. One retry cycle pushes true cost toward $0.55. Against a $2 price, gross margin sits between 70 and 80 percent — until a task needs three retries, at which point it compresses below 40 percent.

That arithmetic matters far more to this industry than it does to the CRM sector. Every onchain agent protocol now raising capital at nine-figure valuations is, whether its founders admit it or not, building a metering business. The two-dollar conversation is the first institutional-scale price discovery for a unit that crypto has spent three years trying to monetize without a reference point.

The technical spine of the offering is not a model. The vendor does not train a foundation model and does not pretend to. The core is a headless toolkit that exposes CRM objects, workflow state, and permission inheritance to external large-language-model frontends over API and the Model Context Protocol — the JSON-RPC specification Anthropic published in November 2024 to standardize how tools and context are handed to a model.

That choice is the entire story. MCP is a wrapper. It is plumbing. Exposing an installed base of enterprise data through a standardized, model-agnostic interface converts a proprietary database into a commodity input, and the vendor has decided that being the commodity input is preferable to being displaced by one. Thirty-seven prebuilt sales skills — meeting prep, deal review, pipeline hygiene — are prompt and workflow templates, not model capability. The data-governance commitment is the substantive engineering claim: customer data neither stored nor used for training. Read that as a compliance boundary, not an architectural one.

The activation figure circulated with the launch — 100,000 users in 35 days — is a distribution metric, not a revenue metric. I have audited enough token launches to hold a hard line on this: activation counts collapse three conversions away from paid usage, and paid usage collapses one more step away from profitable usage. Anyone quoting 100,000 as evidence of product-market fit is quoting a funnel entrance as if it were an exit.

One further caution on sourcing. The primary document I worked from arrived on a Web3 news feed and contained no blockchain content of any kind. Its brand names — an umbrella label, a Claude-branded variant, a Slack-branded variant — do not match any product nomenclature I can independently verify, though the pricing, the protocol, and the model partnership do. Treat the structure as real and the labels as contaminated. This cycle has a chronic problem: aggregation pipelines rewriting press releases, re-dating them, and stripping the provenance out. I have seen the same failure mode in token distribution audits, where a whitepaper's decimals disagreed with the deployed contract's decimals. The document's numbers were right. Its labels were not.

The crypto-side context is symmetrical. Payment-intent standards that let an agent settle a request in stablecoins over ordinary HTTP response codes went from specification to production in under a year. Agent identity registries are in draft. Enterprise budgets are the largest pool of money that has not yet touched a blockchain, and agents are the first plausible pipe between the two.

The metering abstraction is a unit of account, not a discount

Flex credits are the most interesting object in the price list and the least discussed. A vendor-issued credit is a private unit of account. It lets the issuer reprice, rebundle, and retire a product line without renegotiating contracts, because the contract is denominated in credits and the credit-to-work conversion is set unilaterally.

That is seigniorage with a support ticket. Any entity issuing the unit in which a recurring obligation is denominated captures three things: the float, the pricing optionality, and the ability to obscure true unit cost from the buyer. Enterprise agent budgets will be denominated in vendor credits or in stablecoins, and the issuer of that unit — not the model provider, not the chain — will capture the most durable margin in the stack.

Which is why the stablecoin layer is the credible neutral rail here. A corporate treasury will not hold a vendor credit as a store of value; it will hold a dollar instrument with a public reserve report and 24/7 settlement. The integration point between enterprise agents and blockchains will not be an agent token. It will be a treasury operation.

The indifference point is engineered, and that should be stated plainly

Five hundred and fifty dollars per user per month divided by two dollars per conversation is 275 conversations. Below 275 agent invocations per seat per month, the subscription is cheaper. Above it, the meter is cheaper. That threshold did not emerge from a market. It was chosen.

I built utilization curves for lending pools in 2020 and learned something that applies directly. The kink in a lending protocol's interest-rate curve — the point where borrowing cost jumps — is a governance parameter dressed as a market signal. It responds to committee decisions, not to supply and demand in any rigorous sense. The same is true of a 275-conversation indifference point. It is a policy instrument. It tells you what the vendor believes about average usage, and it tells you the vendor expects heavy users to migrate to the meter, where revenue scales with deployment but so does the inference bill.

Reliability is the unpriced risk, and attestation is the missing product

The published material describes permissions and governance settings automatically extending to the agent, with no new model to learn. I have implemented permission inheritance across systems. Perfect cross-object inheritance under delegated authority is not a configuration toggle; it is an unsolved problem in the general case, and it fails silently.

Here is where the onchain side has something genuinely additive to offer, and it is not payments. When an agent updates a pipeline record, drafts customer correspondence, or flags a deal as at risk, the economic consequence is real and the audit trail is a log file controlled by the vendor. That is an invoice without a receipt.

I spent part of 2026 building zero-knowledge attestations that bind an agent's identity to the provenance of the data it acted on — a proof of origin, in the loose sense the phrase is used now. The engineering constraint is cost, and it is not close. A general-purpose proof per agent action at high frequency is economically impossible today. A Groth16-family proof of a moderately complex statement costs cents to low dollars depending on circuit size; recursive aggregation amortizes that to a fraction of a cent per action if you accept a proving latency budget of seconds to minutes. The rule I use is simple: if the proof costs more than the marginal metering fee it certifies, the attestation layer does not exist. At a $2 conversation, a proof budget under $0.30 is viable. At $3, the product is dead on arrival.

So the practical designs are optimistic: bonded challenge windows, sampled verification, fraud proofs with slashing. They are ugly, they are probabilistic, and they are the only things that will ship before 2027.

Settlement demand will reprice blob space, and the models assume otherwise

Agent workloads produce a specific transaction shape: high frequency, small value, heavy metadata. That is exactly the profile layer-2 rollups were built to absorb, and exactly the profile that will consume data-availability capacity faster than any human-driven application.

Blob space is cheap right now. Post-Dencun, the fee market for blobs is separate and, at current utilization, priced near the floor. I have modeled posting rates against projected rollup demand, and the saturation point is a twenty-four-month question at current growth, not a ten-year one. When utilization crosses the target, the fee market does what fee markets do, and per-transaction rollup costs re-rate upward — on the order of a doubling for the rollups with the thinnest margins. Any metering architecture that assumes permanently cheap L2 settlement is underwriting a subsidy with a published expiry date, and agent traffic is the demand that will consume it.

The cycle correlation is weaker than the sector believes

In 2020 I built the correlation table that became the basis of my Liquidity-Cycle Matrix: global M2 expansion against onchain volume, with stablecoin issuance as the intermediate variable. The relationship held well through the retail-driven phase of the cycle, and it explains why the 2021 peak looked the way it did.

Agent workloads break that correlation, and this is the part the token market has not internalized. An agent invocation is triggered by a business process, not by a price chart. The budget line funding it is enterprise operating expenditure, and that budget scales with headcount economics and automation mandates, not with liquidity conditions. Agent flow is therefore a-cyclical or mildly counter-cyclical revenue for a settlement layer — a genuinely more attractive revenue base than retail trading fees, and one that will not correlate with the token price that funded the layer's construction.

Stablecoin mechanics reinforce this. My 2020 stress work on peg stability identified corporate treasury balances as the most peg-supportive demand class precisely because of low velocity — money that arrives and sits. Enterprise agent float is that demand class at larger scale, which strengthens the large issuers and concentrates the market further.

What the four tiers actually hedge

Per-seat revenue is exposed to the very thing the product does. If the agent performs the work, seats shrink. The metered track is insurance against the vendor's own product succeeding. That is not a criticism; it is the most rational pricing architecture available to a company whose installed base is the asset being automated.

The same structural exposure sits inside every onchain protocol selling agent infrastructure today. Their models assume agent transaction counts rise monotonically. Agents net and batch. The assumption will not hold.

The second-order effect is an implementation economy

Total cost of ownership figures circulating for a 30-seat deployment land in the low-to-mid six figures once implementation and credit consumption are counted. That is $6,700 to $15,000 per seat per year. Implementation-heavy pricing recreates the systems-integrator economy in a new dress.

Onchain, the equivalent gap is tooling: budget guardrails, spend caps, agent observability, chain-of-action reconstruction. Whoever builds those becomes the integrator of the agent economy. And a corporate-finance function that did not previously exist now must: a metering-control desk for a line item that scales with machine activity rather than human onboarding.

The consensus framework in this sector holds that agents will pay each other in tokens, that this creates a trillion-dollar payments rail, and that the chains capturing that flow capture the cycle. I do not hold that view, and the enterprise price list is the evidence against it.

Payments are a commodity. A stablecoin transfer over an HTTP response code costs fractions of a cent and settles in seconds; the switching cost between two dollar-denominated rails is approximately zero. A token that exists to mediate agent payments must therefore justify a premium for a service that already has commodity providers. It will not.

The scarce inputs are elsewhere: metering accuracy, verifiable attribution of what an agent actually did, budget enforcement that stops a runaway loop from generating a five-figure invoice, observability into multi-step chains. None of these are payment rails. All of them are attestation and control-plane products.

The decoupling thesis, stated precisely: onchain agent activity will grow while the fee revenue and token value accruing to the rails that host it decouple from that growth. Two mechanisms drive the wedge. Agents batch and net, and netting collapses transaction counts relative to economic volume, so block-space demand understates agent throughput by an order of magnitude. Enterprise deployments will also run on permissioned or hybrid execution environments with near-zero public fee capture, because data residency and audit obligations make fully public settlement a compliance problem rather than a scaling problem.

The jurisdiction competing hardest for those deployments is not doing so out of enthusiasm for the technology. It is replicating a licensing playbook already run once in this region against a neighbouring financial centre, applied to a new asset class with a different risk register.

The bull-market version of this narrative will not survive a single quarter of published agent task-success rates. Which is why nobody publishes them.

The signals worth tracking are not price signals. Whether any vendor publishes an end-to-end task success rate for multi-step enterprise workflows, with a rollback mechanism described. Whether metering attestations become verifiable by a third party rather than merely asserted. Whether the implementation cost curve falls fast enough that mid-market firms are not structurally excluded from the agent economy — because an automation layer only the top decile can afford is not a productivity revolution, it is a margin transfer.

The question that resolves this cycle is not which chain agents run on. It is who signs the receipt. Exit strategies are written in ice, not in hope.