Monday.com's new AI credits are a token system without a token. The market saw the 12.6% stock pop after the pivot announcement and called it a fresh narrative. I saw a centralized metering system with a trust problem. Math doesn’t negotiate. Neither does the cost of running an AI model.
Over the past year, the enterprise collaboration giant has transformed its product from a "Work OS" — a place where teams organize tasks — into an "AI Work Platform" that actually performs work. The shift sounds like branding. It is not. It is a structural migration from subscription-based software to consumption-based utility. And it carries a technical and economic DNA that DeFi protocols have been wrestling with for years: metering, settlement, marginal cost leakage, and the trust assumptions hidden inside a resource economy.
This is not a story about Monday.com. This is a story about how blockchain's core problems — verifiable measurement, transparent settlement, and trustless execution — are about to become the defining problems of enterprise AI.
The Metering Stack Nobody Mentions
Monday.com's pricing structure, established in May 2026, is deceptively simple. Basic plans include 1,000 credits per month. Standard gets 2,000. Pro gets 3,000. Overspend and you pay $0.01 to $0.0125 per credit, depending on whether you commit annually or monthly. Annual prepayment gets a 25% discount. That's a classic prepaid discount model designed to smooth cash flow and lock in commitments.
Under this surface lies an infrastructure problem. To make AI credits work, Monday.com must run a real-time resource metering engine that tracks token consumption, compute cycles, tool invocations, data throughput, and per-tenant agent activity. It must map those heterogeneous usage events into a single fungible unit — the credit — and bill customers for it. This is not a feature. This is the construction of a lightweight cloud billing platform.
I've built similar systems. During the 2022 bear market, I spent six months writing a minimal zkSNARK prover in Rust to understand Groth16 constraints. That experience taught me how hard it is to separate computation cost from proof generation. Enterprise metering is the inverse of that problem: instead of proving computation happened, you must attribute every cost to a customer. Monday.com's engineers are now living inside that complexity.
The technical burden is real. A multi-model abstraction layer must talk to Anthropic, OpenAI, and Microsoft without leaking data or introducing inconsistent behavior. An agent runtime must orchestrate workflows, maintain state across asynchronous calls, and recover from partial failures. And the metering engine must do all of this with sub-millisecond accuracy. Code is law, but bugs are reality. A bug in the metering engine means revenue leakage. A bug in the agent runtime means data leakage. Both are existential for a public SaaS company.
This is why the 20% layoff — 620 to 630 people — makes tactical sense. The company needs fewer enterprise relationship managers and more distributed-systems engineers. The reorganization charges of $45 to $55 million are the cost of buying back its own architecture stack.
The Margin Squeeze Nobody Is Calculating
Traditional SaaS enjoys marginal service costs that approach zero. Software delivery is a pure digital copy. AI credits are different. Every credit burned triggers a real API call to Anthropic or OpenAI. The marginal cost of satisfying a credit is external and non-negotiable. Model providers set the price. Monday.com merely resells that cost with a markup.
Assume the markup is 2x to 3x. If Monday.com charges $0.01 per credit, the underlying model cost might be $0.003 to $0.005. That leaves a gross margin of 50% to 70% on consumption revenue. Compare that to the 75%–85% gross margin of traditional SaaS. The more AI credits a customer consumes, the more the blended gross margin deteriorates.
This is a hidden unit economic problem. In blockchain terms, it is the difference between a protocol that charges transaction fees to cover validator costs and an L2 that pays a separate data-availability provider on the base chain. The external cost doesn't go away. It just gets passed through.
Monday.com's CEO reaffirmed 19–20% revenue growth guidance. But that guidance has not been stress-tested against the margin drag of AI consumption. The real number investors should watch is not revenue growth — it's the take rate. What percentage of each AI credit dollar remains after the model provider takes its cut? If that number is below 40%, the AI transformation becomes a volume trap.
The problem is not unique to Monday.com. Every SaaS company that resells third-party AI models faces the same tension. The solution — building proprietary models — is extraordinarily expensive and outside Monday.com's stated strategy. They have chosen the multi-model connector path. That's a rational decision, but it caps their long-term margin ceiling.
ARR Quality and the Consumption Fallacy
Here is the question that should keep public-market investors up at night: does AI credit prepayment count as annual recurring revenue? In the classic SaaS model, revenue is recognized ratably over the subscription term. Consumption revenue is different. Customers buy a bucket of credits. They may or may not use all of them. Unused credits are a liability, not an asset.
If Monday.com records credit prepayments as ARR, the reported growth rate overstates the underlying recurring revenue. The company needs to disclose three things: the timing of revenue recognition, the liability for unconsumed credits, and the split between seat-based and usage-based revenue. Without this disclosure, the market is pricing a blind box.
The same problem exists in crypto. Token-based protocols often report "protocol revenue" as the value of fees paid in tokens, but if those tokens are funded by a treasury rather than real usage, the number is fiction. I saw this pattern in the LUNA collapse. In 2021, I spent three weeks dissecting Anchor Protocol's smart contracts. The death spiral wasn't an oracle bug at first glance — it was a structural misalignment between high fixed yields and the borrowing demand needed to sustain them. The protocol reported revenue that depended on assumptions that were unsustainable. Code is law, but bugs are reality. The bug was in the economic model, not the Solidity.
Monday.com's AI credit model has a similar structural risk. If customers adopt AI agents for only low-risk tasks, they will buy 1,000 credits, use 500, and never renew. The renewal rate will be based on value delivered, not seats filled. In traditional SaaS, revenue is sticky because users are hard to remove. In consumption models, revenue is sticky only if the AI produces outcomes that justify ongoing spend. That's a much higher bar.
The Efficiency Paradox and the Oracle Problem
The deeper issue is what I call the AI efficiency paradox. As language models improve, they become more token-efficient. The same task that required 1,000 tokens two years ago might require 300 tokens next year. For a customer paying per credit, that means the same outcomes cost fewer credits. Revenue per customer shrinks even as customer satisfaction rises.
This inverts the traditional SaaS flywheel. In classic SaaS, product improvements increase retention and expansion. In metered AI, product improvements reduce consumption and therefore reduce revenue. The only way to counteract this is to anchor pricing to business outcomes rather than model operations — charge per successful task, not per inference. But outcome-based pricing is hard to verify. How does the customer know the task actually succeeded? How does the vendor prove it did not inflate the outcome metrics?
This is precisely the oracle problem that blockchains have been wrestling with for years. Oracles are trusted data feeds that connect external truths to smart contracts. In AI, we need a different kind of oracle: one that independently verifies the execution of an AI agent and its cost. Zero-knowledge proofs could be the answer. A ZK circuit can prove that a specific model ran on a specific input and produced a specific output, without revealing the input or the intermediate state. That would enable verifiable inference — a cryptographic receipt for AI execution.
Privacy is a feature, not a bug. In 2025, I collaborated with a legal-tech startup to build a ZK-proof circuit for a DeFi lending protocol that verified creditworthiness without exposing personal data. We optimized proving time from 500ms to 150ms. But that was a static verification of a known set of inputs. AI agents are dynamic and stateful. Verifying them cryptographically is a different magnitude of problem.
The enterprise implications are massive. A company sending its workflow data to Anthropic's API has no audit trail. Did the model train on that data? Is the data stored? What are the retention policies? Monday.com is asking enterprise customers to trust that their process data remains private inside a third-party model's black box. And the company itself is a conduit, not a controller. The threat surface is not Monday.com's own infrastructure — it's the supply chain of the model providers.
This is the blind spot of the AI transformation. The competitive moat is not the number of AI models connected. It's the ability to guarantee data confidentiality, usage verifiability, and cost transparency. Without cryptographic assurance, enterprise clients will only delegate low-stakes tasks to AI agents. Credit consumption will stay shallow, and the new revenue model will underperform.
The Real Threat Is Upstream, Not Lateral
It's tempting to compare Monday.com to Asana, ClickUp, or Notion. But the existential competitor is Microsoft. Monday.com connects to Microsoft's AI models while simultaneously competing with Microsoft Teams and Project. That's a strange position: being a customer of your largest competitor's infrastructure and hoping they don't squeeze you.
OpenAI and Anthropic are also moving downward. If they ship enterprise agent orchestration layers — and they will — the "work OS" becomes a thin commodity. The multi-model connector strategy is a Trojan horse. It gives Monday.com AI capability quickly, but it also hands over the data flywheel to the model providers. The more workflows run through because of the connector, the more the model providers learn about enterprise work patterns. They can then package that intelligence into their own products.
In DeFi, we call this vertical integration. A chain that builds a top-level aggregator ends up competing with the protocols it aggregates. Monday.com risks being both a distribution channel and a collection layer for Microsoft's large enterprise graph. This is not a sustainable equilibrium.
What would a defensive posture look like? Monday.com could build proprietary model fine-tunes on its own workflow data, so that the value-add is not the model itself but the context-specific behavior. Or it could adopt privacy-preserving inference technologies — local models with ZK proofs, or at least zero-retention agreements with model providers. But the latter is not legally enforceable in a meaningful way. The only enforceable guarantee is cryptographic.
The 25% premium for monthly billing suggests Monday.com is already thinking like an infrastructure provider — encouraging upfront payments to reduce collection risk. But the bigger risk is being upstreamed by the companies it depends on. The market has priced Monday.com as a growth story again. I see a company running a centralized token system with a counterparty risk that no token alone can solve.
Contrarian: The Real Value Isn't AI — It's the Audit Trail
The contrarian take is not that Monday.com will fail. It's that the AI work platform will eventually be judged on something none of the marketing materials mention: auditability. The companies adopting AI agents will face board scrutiny, insurance requirements, and regulatory examination. They will need to prove what an AI agent did, when it did it, and why. This is the digital ledger problem. Blockchain technology is a natural fit for this, not because of token incentives, but because of its immutable, timestamped, and independently verifiable structure.
Monday.com could become a big data capture layer for human work. But AI agents produce a different kind of trace: tool calls, model inferences, and state transitions. To make that trace trustworthy, you need cryptographic primitives — Merkle trees for state, digital signatures for agent identity, and ZK proofs for privacy-preserving verification. Monday.com is not building any of that. The AI credits are merely a financial instrument, not a proof system.
The blind spot is that the true asset of an AI work platform is not its feature set; it's the legitimacy of its execution history. I audited institutional custodians in 2024 after the Bitcoin ETF approval. The gap between their marketing and their actual multi-signature thresholds was glaring. The point is not that they were malicious; it's that they could not prove they were secure. The same gap exists here. Monday.com cannot prove what its AI agents actually did. The data lives in foreign model providers' logs.
That's the opportunity blockchain presents. A chain-based agent runtime could emit a receipt for every action: hash of input, model version, output, and the resulting state change. The credits could be tokens on a ledger, with transparent issuance and burn. The AI pricing model would become self-auditing.
Takeaway
In the next 18 months, every major SaaS vendor will copy Monday.com's credit model. They will all discover the same reality: metering is easy; trust is hard. The enterprise will not surrender its most sensitive processes to a black box without an audit trail. Some startup will build a "verifiable inference" standard that wraps AI execution in cryptographic proofs, and the CREDITS will become tokens. Math doesn’t negotiate. The market will eventually demand proof.
Privacy is a feature, not a bug. But proof is the feature that makes privacy possible. The question is not whether Monday.com survives. The question is whether the AI work platform category will eventually be co-opted by protocols that treat computation as law. The clock is ticking on the centralized meter.