The AI Credit Ledger: Centralized Sequencers, Metered Work, and the Monday.com Token That Isn’t

In May 2026, Monday.com crossed a line that most enterprise software firms are afraid to name: it stopped selling a tool and started selling a metered resource. For a decade it had been a Work OS, a canvas where teams arranged tasks and talked about progress. With the pivot to the AI Work Platform, the canvas became an actor. Seats became access. Workflows became agents. And the unit of exchange became an “AI credit” — 1,000, 2,000, or 3,000 minted into each tier of subscription, redeemable at roughly one cent apiece once the monthly balance runs dry. The announcement landed after a brutal year: the stock had fallen more than 50 percent since January, roughly 620 to 630 people — about 20 percent of the workforce — had been cut, and the company had absorbed 45 to 55 million dollars in restructuring charges. The market’s response on announcement day was a 12.6 percent bounce. I have learned to distrust bounce patterns. They usually mean the crowd has found a comforting story to replace the numbers it no longer understands, and sometimes the story is true for a quarter and false for a decade. The question is whether this particular story has a soul in it, or just a billing engine in a trench coat. Here is what the crowd understood. Monday.com’s 225,000-plus customer organizations gave it a distribution base that most AI startups can only hallucinate. Its old business had the cleanest economics in software: a seat-based subscription with near-zero marginal delivery cost, gross margins between 75 and 85 percent, and renewals as predictable as payroll. The new business is a different animal entirely. The AI Work Platform connects natively to Anthropic, OpenAI, and Microsoft models, wraps them in one-click connectors, and lets non-technical team members configure autonomous agents that actually execute work. That is a genuine evolution from recording system to action system. But the commercial machinery underneath it is the thing nobody discussed on the earnings call: a metering and billing engine that tracks every agent’s token consumption, tool calls, data throughput, and compute resources, and maps them to a billable credit. I built enough observability infrastructure in my data-science years to know what that stack costs. It is not a feature. It is a lightweight cloud billing platform, the equivalent of building a settlement layer inside a single company instead of across a network. The layoffs were not a cost-cutting exercise. They were the invoice for that engine. The structural risk lives upstream, in what I have come to call the three pools problem. In the years after the fourth Bitcoin halving, I watched miner revenue collapse and hash power coalesce around a small number of dominant pools; the protocol remained technically open while consensus became operationally concentrated. The same shape is visible here. Monday.com connects to Anthropic, OpenAI, and Microsoft, but that is not diversification; it is delegation to a cartel of three validators. Each one of those validators controls the cost basis of every credit minted, the latency of every agent action, and the policy terms of every piece of enterprise data that crosses the model boundary. The multi-model connector is a beautiful abstraction layer, but abstraction layers do not create neutrality. They create the illusion of optionality while the real pricing power accrues to the people who own the weights. It is the decentralized sequencing promise all over again: two years of PowerPoint slides, one live node. I have audited enough oracle designs to know that a middleware that merely routes requests is a thin moat: beautiful until the upstream validator decides to build its own front end. The crypto analogy here is not decorative. An AI credit is a claim on an action, minted by a single authority, settled by a closed-loop accounting engine that nobody outside the company can audit. If you squint, it is a centralized stablecoin: pegged to the promise of “one actionable workflow,” backed by the company’s contracts with three external model providers, and redeemable only within the issuer’s own walled garden. The reserve quality question, the one that haunted algorithmic stablecoins, becomes a margin question. Traditional SaaS gross margins run near 80 percent. If model API costs consume between 30 and 60 percent of every credit dollar, the blended margin sags toward the high 50s or low 60s. And then the company’s core unit economics become a function of OpenAI’s pricing sheet. This is the same maturity mismatch we audited in DeFi yield products like sUSDe during the bull market: clever architecture that works while the underlying assumption — rising consumption, falling model costs, sticky credit prices — holds, and that unravels in the quarter when all three assumptions get tested at once. The 25 percent premium on month-to-month credit purchases, compared with annual prepayment, is a standard prepaid discount. But it is also a tell: the company wants its customers to prefund the meter, to stabilize the treasury, to deposit locked liquidity into a vault whose liabilities are denominated in somebody else’s API. The data flywheel is the quiet asset in this story, and it is a double-edged sword. Every agent execution on Monday.com generates a behavioral record: which workflows get automated, how users correct AI mistakes, which prompts fail and which succeed. That corpus is a training goldmine, and the company’s 225,000 customers give it a breadth no AI-native startup can match. But the same corpus is a liability. Enterprise clients are terrified of sending their workflow data to external models; the legal and procurement teams will demand zero-retention agreements or on-premise deployments before any sensitive process is automated. The practical consequence is that only low-risk tasks will be handed to agents in the first year. And if the agents are only trusted with trivial work, credit consumption will stay thin, and the metered revenue engine will sputter long before it flies. The architecture is sound; the trust layer is missing, and trust does not ship in a release train. Now layer in the revenue recognition question, because this is where the market narrative is most fragile. Metered consumption revenue is not subscription revenue. If prepaid credits are booked as ARR at purchase rather than at consumption, the recurring number is diluted by liabilities dressed as growth. Unused credits become deferred revenue, a stack of unearned promises that behaves exactly like a token treasury revalued each quarter. Investors rewarding the 12.6 percent bounce are betting that the AI transformation story outweighs the contamination of the SaaS model’s most sacred metric. Maybe they are right. But the same logic once sustained the thesis that NRR above 120 percent could ignore the fact that growth came from inflating collateral rather than serving users. The AI efficiency paradox is the next stake in the heart: the better the model, the fewer credits a customer needs for the same outcome. In the seat world, revenue scales with headcount and enthusiasm. In the metered world, every improvement in underlying intelligence is a headwind to usage. This is the opposite of the old SaaS flywheel. It is a flywheel that spins backward, and the only way I see to fix it is to anchor pricing to business outcomes rather than inference inputs. But outcomes are hard to verify — which is precisely the moment you want a public, tamper-evident record of what the machine did. There is also a quieter operational cost that the 12.6 percent bounce conveniently ignored: the sales motion has fundamentally changed. Per-seat pricing was a one-line conversation; the buyer asked how many people, the vendor quoted a number, and the deal closed in days. Per-credit pricing requires the seller to explain how many credits an agent consumes while performing a hundred tasks, what a credit is worth in business terms, and why the monthly rate costs 25 percent more than the annual rate. That is value selling, and value selling is expensive. Sales cycles stretch from weeks to months, customer education budgets balloon, and customer success stops being software training and becomes AI process consulting. In my experience building decentralized protocol tooling, every time we added a metering concept to the user experience, acquisition costs roughly doubled and churn tripled among users who did not understand the unit of account. Monday.com is about to relearn that lesson at enterprise scale. Switching cost is the third pillar of the new architecture, and I want to be honest about its double edge. Configuring thirty AI agents to handle customer service workflows creates a binding far deeper than legacy SaaS data migration. The agents encode a company’s behavioral logic: its tool calls, its approval thresholds, its dark corners of exception handling. Exporting that to a competitor means re-writing the DNA of the operation. I called this lock-in “liquidity stickiness” in my protocol PM days and treated it as a bull case. In reflection, the bull case was propaganda. When the lock-in sits on a private ledger owned by one corporation, it is not decentralization. It is a concentrated validator set with a captive user base. The internal dynamics are equally telling. Once IT departments buy AI credits, they need to allocate monthly stipends to marketing, finance, and operations teams, monitor consumption, and argue about who overspent. You have just imported the entire token economics discipline — vesting, budgets, sybil resistance against departments gaming their allocation — inside the enterprise, with a single administrator holding the governance keys. Cloud FinOps teams have played this game with compute; now they will play it with agent souls. We chart the code, but the soul chooses the path. The Microsoft question deserves special attention because it is the quiet engine of the entire arrangement. Monday.com is simultaneously a customer of Microsoft’s AI infrastructure, a connector to OpenAI’s models, and a direct competitor to Microsoft Teams and Project. That dual identity is institutionalized dependency. The moment Microsoft decides to ship an enterprise agent orchestration layer inside its own ecosystem, the middleman value of the connector evaporates. I watched the same dynamic destroy middleware protocols in the last cycle: the oracle that aggregates upstream data is always one vertical integration away from obsolescence. The hedge is obvious — deepen the Anthropic relationship, build proprietary agent memory, own the workflow graph — but the hedge is also expensive, and it competes with the very margins the model API costs are already compressing. The contrarian view, and I mean genuinely contrarian rather than soothing, is that this pivot is the most honest thing enterprise software has done in a decade. Consumption pricing forces the vendor to admit that value lives in action, not in attendance. An agent that never completes anything mints no credit. The meter is, economically speaking, a proof-of-work protocol: energy — in the form of model inference — must be spent to produce value, and the user pays for the energy rather than for the privilege of sitting in the office. That is closer to the ethos of verifiable work than the old subscription rent. The tragedy is not the metering. The tragedy is the closed ledger. There is no customer-accessible way to audit how many credits a given invocation consumed, when the meter started, or what precisely the agent did with the data. The workers whose time and judgment trained these workflows have no record of the arrangement. I spent 2022 auditing failing L1 protocols, identifying three centralization vulnerabilities in their consensus models; the same checklist applies here: single sequencer, opaque oracle, and no exit-settlement mechanism. We chart the code, but the soul chooses the path — and in this architecture the path is a permissions file owned by someone else’s database. The 19 to 20 percent growth guidance will be met or missed; the stock will rally and fade. The ledger question will not. If consumption-based intelligence is the future, the future demands an open record of consumption: portable agent definitions, verifiable usage proofs, and the right to exit without rewriting your operational memory. Monday.com built a beautiful meter. The rest of us should be building the block explorer for it. Because the credit counter slowly accumulating on an employee’s screen is not just a pricing mechanism. It is a measurement of human time, compressed into units, denominated in a currency the worker cannot verify and cannot carry out the door. Protocol neutrality was always a myth, but the comfort of paying one vendor for the privilege of being measured is a worse one. The market’s 12.6 percent cheer was hope, not proof. If the AI credit economy deserves its name, the credits must be auditable and the paths must remain sovereign. We chart the code, but the soul chooses the path. Let the meter run — but only if the ledger stays open.