The Metering Layer Is the New Mint: Enterprise AI Credits Are Rebuilding the ICO's Worst Failure

Two numbers define the entire enterprise agent market, and neither vendor will put them side by side. Salesforce prices a Flex Credit at $0.005. Microsoft prices a Copilot Credit at $0.008. That is a 60 percent spread on what both companies call "a credit." Neither unit converts to the other. Neither converts to dollars at a locked rate. Neither carries a burn-rate disclosure. And each business action burns a different quantity: Salesforce pins a refund, an order lookup, and a customer notification at 20 credits apiece — roughly $0.10 — while Microsoft refuses to define the unit at all. I have read this contract before. In 2017 I sat in a Tokyo office with forty token white papers and audited every one against a fifty-point hygiene checklist. The failure mode was never the token itself. It was the metering. When the issuer alone defines what a unit is worth, the unit is not a price. It is a promise with a decimal point. Chaos demands structure before it yields value. Enterprise AI just walked into the same room. For fifteen years software was billed by the seat: a fixed price per user per month, multiplied by headcount, discounted by volume. That model held because usage was bounded. A word processor does not cost more to run when the typist types faster. An autonomous agent breaks the assumption at the physical level. Every tool call, every retrieved document, every injected turn of history enters the context window and consumes tokens. The seat was a unit of access. The agent needs a unit of work, and the industry has not agreed on what work is — which is why the invoice has become unreadable. Eight vendors are improvising answers, and they disagree on the fundamentals. Microsoft and Salesforce publish rate cards but bury the real variables — credit burn rates, default configurations, SKU lifecycles — inside partner portals. Anthropic publishes a $20 seat and bills usage separately at API rates, so the seat is an entry ticket, not a price. OpenAI routes enterprise buyers to a sales form with no self-serve number. Google and ServiceNow publish nothing; third-party estimates for Google float between $21 and $60 per seat and have never been confirmed on an official page. xAI and Meta sit at the opposite pole, publishing API cards openly because transparency is their only acquisition lever. That split is not random. It maps precisely onto market power. The stronger the vendor, the less it shows you. Opacity is not a shared industry flaw. It is a moat, and the moat is widening. This matters beyond AI, because the same question is about to land on every DAO treasury that pays for autonomous execution: if you cannot audit the meter, you cannot budget the spend. Start with the arithmetic, because the arithmetic is where the marketing dies. Meta publishes $2.00 per million tokens for WhatsApp replies. Divide that into a typical per-message cost of four to five cents and you get 20,000 to 25,000 tokens consumed per message. That is ten to fifty times a normal chat turn, which runs 500 to 2,000 tokens. The excess is not waste. It is the architecture: system prompt, tool definitions, conversation history, retrieval results, and tool return payloads all enter the window on every cycle. An agent does not answer a question. It assembles a context, calls a tool, reads the result, and reassembles. Every step is billable, and none of it appears on the invoice. Microsoft admitted the consequence in its own deprecation notice, retiring the $200-for-25,000-messages SKU on the grounds that consumption depends on how the agent is designed and used, so a message no longer maps to a fixed price. That is a vendor conceding the same word can mean ten different costs. Salesforce pushed the abstraction up a level and metered by "action" at 20 credits each, but an action is just as elastic: a status lookup and a refund-plus-notification are both one action and nowhere near the same cost. Anthropic meters by token at API rates yet sells the seat as if the seat were the product. Three metering units, three risk allocations. Bill by token and the supplier carries the efficiency risk, because a customer who writes better prompts pays less. Bill by action and the customer carries the complexity risk, because a heavy task compressed into one action favors the buyer while a light task split into three favors the seller. Bill by conversation and both sides gamble on a number neither controls. The choice of metering unit is not a billing detail. It is the pricing decision itself. Whoever defines "one action" can raise the effective price without touching the sticker. This is where my 2017 audit scars activate. A credit is a private currency issued by a single vendor, and it carries three properties I have flagged in every token review since: the issuer can change the burn rate unilaterally, the unit is non-convertible and non-transferable, and there is no disclosure obligation when the issuer debases it. Salesforce can move 20 credits per action to 30 next quarter while the rate card still reads $0.005. Microsoft can redefine what a Copilot Credit covers and the price per credit never moves. A unit whose burn rate is set by the issuer is not a unit of account. It is a claim on the issuer's future discretion — structurally identical to the non-dividend governance tokens I spent a decade warning clients about, value that exists only if a later buyer accepts the same undefined unit. Now watch the default. Microsoft is migrating Copilot Business through its cloud partner channel to usage-based billing with metering switched on by default — opt-out, not opt-in — at a default ceiling of 4,000 credits per user per month, with the United States not on the exclusion list. Microsoft's own worked example is 100 users capped at 400,000 credits monthly. Run the Copilot Studio rate of $0.008 per credit against that ceiling, purely as an upper-bound analogy, and you get roughly $3,200 a month, or $32 per user per month of potential overage stacked on the seat. The number is an estimate. The behavior is not. Default effects convert at brutal rates because most procurement teams never revisit a setting they did not consciously choose. A billing mode that activates without a decision is not a pricing model. It is a consent bypass. Then there is the archaeology problem. Salesforce still displays a $2-per-conversation card, annotated for existing Agentforce Conversations customers only. Buyers searching for prices, analysts quoting third-party reviews, and archived screenshots will all surface a number that may already be dead. Compare it honestly: if a conversation averages 20 actions, the two models are equivalent; if it averages five, action billing is four times more expensive. That conversion ratio — the one number that decides whether an existing customer's bill quadruples — has never been published. Price archaeology is a trap, and it is set deliberately. Strip the marketing and four typologies remain. Quote-gated vendors — ServiceNow, Google, xAI enterprise, OpenAI — practice first-degree price discrimination and quietly exclude every buyer below an undisclosed minimum commit. Rate-card vendors — Microsoft, Salesforce — publish an anchor and hide the variables, so you can estimate but never lock. Seat-plus-API vendors like Anthropic publish a number that badly understates total cost. Only Meta's WhatsApp rate and xAI's public API card support honest unit economics, and both are acquisition instruments, not market prices. The buyer comparing these four is not comparing prices. He is comparing four different definitions of a dollar. Read the pattern and the real story surfaces: value capture is migrating from the seat to the metering layer, and the metering layer is the new control point. Cloud computing taught the same lesson. AWS's moat was never EC2; it was CloudWatch, the billing system, and reserved-instance pricing. Whoever defines the meter defines the standard. Salesforce pushes credits, Microsoft pushes credits, Meta and xAI push tokens, and that is a standards war wearing the costume of a transparency complaint. Why do the strongest vendors hide the price? Because publishing a number they may have to raise hurts trust more than publishing nothing. Agent token consumption erodes gross margin far faster than chat did — a single multi-tool task can burn the equivalent of a hundred chat turns — and a vendor who posts a per-action rate that doubles in a quarter has handed every competitor a talking point. Silence preserves optionality. It also preserves the ability to price the same workload differently for two customers on the same day. The channel change proves it. Microsoft is simultaneously unbinding the base subscription, swapping SKU identifiers, and switching default metering, and it already delayed the date from November 2 to December 1. The delay is a signal: the partner channel is not ready. Small managed service providers lack the billing systems to meter usage at all, so SKU churn and add-on renaming become back-office re-engineering costs that roll downhill to the end client. And the reference deployment Microsoft cites — 30,000 seats, cost undisclosed — is measured on adoption, not productivity. The KPI is how much was consumed, not how much was saved. Under an adoption KPI, metered billing rewards overconsumption by design, the exact opposite of what the buyer wants. There is a regulatory shadow the vendors have not priced. If Microsoft bundles Copilot consumption with Azure spend — its own documentation ties Azure subscription setup to the license — and switches metering on by default, that is textbook tying and self-preferencing, and European competition authorities have long memories. A pricing model that would draw scrutiny as a contract clause is being shipped as a default setting. The missing primitive is attestation. No vendor currently lets a buyer verify that a given action actually consumed 20 credits. There is no signed receipt, no reconciliation log, no third-party audit. Every invoice is a take-it-or-leave-it statement. Last year I helped three protocols build a verifiable-credential system so autonomous agents could transact against decentralized exchanges with accountability, and the lesson carries straight over: identity and metering are the same problem. An agent that cannot prove what it did cannot be priced, audited, or held liable. Identity without utility is just noise, and a meter without an audit trail is just a claim. For the crypto-native reader the irony is sharp. We spent a decade building the exact tooling enterprise AI now needs — signed usage attestations, Merkle-anchored logs, programmable settlement — and almost none of it has crossed into these rate cards, even though the firms building agent-payment rails are the same firms selling the opaque meters. We debated verifiable compute for years while the SaaS giants quietly shipped discretionary metering and called it a roadmap. Underneath all of it sits the cost nobody meters: the human who reviews what the agent did. A refund, an order change, a customer email fired at $0.10 an action still needs a person to confirm it was correct before it ships. That labor is real, recurring, and absent from every vendor invoice, every ROI spreadsheet, and every headcount line labeled AI. It is absorbed into existing roles and reported as a productivity gain. Human review is the load-bearing wall of every agent deployment, and the entire pricing structure assigns it a value of zero. Here is the pragmatism test, and it cuts against the comfortable narrative. The instinct is to demand transparency and assume it solves the problem. It does not, for two reasons. First, opacity is not a market failure — it is the mechanism by which incumbents protect margin and avoid price wars, exactly as designed. The vendors who publish clean rate cards are the ones with the least to lose, and demanding that Microsoft match xAI's transparency is asking the leader to surrender its strongest weapon. Second, on-chain metering is not automatically honest. I spent 2020 mapping DeFi yields into risk matrices and watched "market-driven" interest models on Aave and Compound turn out to be arbitrary curves dressed in math. Putting a meter on a chain does not make it truthful; it only makes the lies permanent and public. And a credit redeemed only by its issuer is not so different from the governance tokens I have dissected for years — an instrument whose only exit is a later buyer who accepts the same undefined unit. Trust is built through transparency, not promises — but transparency of a discretionary rate is theater, not proof. The buyer who accepts a credit at face value is not hedging the risk. He is holding it. So do not ask these vendors for a price. Ask for the meter. Demand the credit burn rate locked in contract, an annual ceiling on the rate, a published action-to-conversation conversion ratio, overage behavior specified down to whether the service halts, unused-credit carryover terms, and audit access to the metering logs. If the vendor will not expose the meter, the meter is the product, and you are not buying software — you are renting access to someone else's discretion. The market that standardizes the metering layer first will set the terms for everyone after it. We do not speculate; we engineer certainty. That work has not started. It is the only work that matters. Until then, the invoice is the only contract that matters, and it is written in a currency you do not control.