The Unpriced Reviewer: Agentic AI's Load-Bearing Cost and the On-Chain Layer That Will Finally Price It

Two numbers landed in the same quarter and nobody reconciled them. Notion told the market its agents cut operating costs by 90%. A Harvard working paper on AI-assisted knowledge work reported that time spent by professionals reviewing AI output rose 49%, and the share of staff dedicated to review rose 14%. Both numbers can be true simultaneously. That is precisely the problem.

While the market reads the first number as a productivity event, the second number is a liability event, and it is not booked anywhere. No vendor income statement carries it. No customer capex line captures it. It sits in payroll, diffuse and unnamed, and it scales with deployment rather than against it. In a bear market, that is where the drawdown lives — not in the narrative collapse, but in the multiple that was built on the assumption that marginal cost goes to zero.

Liquidity doesn't price what it cannot see. And right now the entire agent economy is running on a cost curve that nobody has marked.

What the Harness Actually Is

Strip the marketing and an AI agent is three components: a model, a tool interface, and a harness. The harness is the part nobody demos. It decides what the model is allowed to touch, in what order, with what rollback. Everything interesting about the current generation of products lives in the harness, not in the model weights.

Two harness designs dominate. The first is a guardrail model: a second model evaluates the proposed action before execution and blocks or allows it. Notion's "Review for me" is this pattern — an independent model checks the agent's action before it commits. The second is human-in-the-loop gating: Microsoft's Guardian Agents framework requires provable human approval at defined checkpoints. One design has AI reviewing AI. The other has a human reviewing AI. They are not variations on a theme; they are different cost structures wearing the same product label.

Satya Nadella said the quiet part out loud: the systems that trust the model least are the most trustworthy. Read that as a strategic disclosure rather than an aphorism. Microsoft is not betting its agent franchise on model capability. It is betting on the review layer. That is a deliberate divergence from the labs, which are betting on capability and letting the harness catch the failures.

Why does the harness exist at all? Because of compounding error, and the arithmetic is unforgiving. A ten-step task with 95% per-step success completes end-to-end 59.9% of the time. A fifty-step task at 99% per step lands at roughly 60.5%. You can improve the model or you can insert a checkpoint. There is no third option, and the checkpoint is cheaper than the accuracy gain at the frontier. Review is not a safety feature bolted onto agents. It is the mechanism that makes long-chain autonomy economically usable at all.

That distinction matters for how the cost is framed. If review is a safety net, it is discretionary and gets cut when budgets tighten. If review is load-bearing, it is structural and behaves like a utility bill. The industry has been marketing it as the first and accounting for it as the second.

The Divergence Nobody Marks

Follow the cost of each party.

The vendor's cost is inference compute. It is predictable, it is deflationary, and it responds to routing optimization — send the trivial steps to a small model, reserve the expensive model for the hard ones. Notion's 90% figure almost certainly comes from routing, not from capability gains. That is a legitimate engineering win. It is also a disclosure that the capability gap did not close; it got covered.

The customer's cost is review labor. It is unpredictable, it is inflationary, and it grows with the autonomy you grant the agent. The more you let the agent do, the more output a human has to check. The two cost curves move in opposite directions. The vendor's line falls while the customer's line rises, and the invoice captures only the first.

That is the mechanical origin of what the source analysis calls a false sense of profitability. It is not fraud. It is a measurement gap with a price attached, and the price is being paid by the party with the least ability to reprice it — the mid-market operator who bought ten seats and discovered that the real deployment cost was a fraction of a full-time reviewer per agent.

Pricing theory explains why this persists. Seat-based and usage-based SaaS pricing can only capture variables the vendor can observe. Review load is unobservable to the vendor, heterogeneous across customers, and correlated with exactly the use cases that generate the most usage revenue. Two firms deploy ten agents each. One uses them for internal drafting with low consequence of error. The other uses them for customer-facing commitments with contractual consequence. The second firm's review burden might be five times the first's. They pay the same subscription.

The information asymmetry is not a side effect of the business model. It is the business model. A vendor that priced review load honestly would publish its own agent's real ROI, and that number is uncomfortable at this stage of the cycle.

The Regulatory Clock Starts Ticking

Colorado's ADMT framework changes the accounting category. The proposed rules require human review within 45 days, conducted by an independent reviewer with actual authority to approve, modify, or overturn the automated decision. Read the requirements literally: a defined deadline, a named responsible party, a documented authority, and a legal consequence for failure.

That converts review from an internal cost center into a compliance obligation. Internal cost centers get optimized, deferred, and quietly under-resourced. Compliance obligations get audited. Once a number has a deadline and an enforcement mechanism, it enters the budget as a line item rather than a residual.

Two caveats, stated plainly. First, these are proposed rules, not settled law; the January 2027 effective date is an intention, not a fact, and the rulemaking can narrow before it lands. Second, the 45-day window is generous on its face and misleading in practice. If the review volume per reviewer is high, the effective time per case collapses toward minutes, which is exactly the condition under which automation bias bites hardest. Reviewers under time pressure defer to machine output. A 45-day clock does not guarantee scrutiny; it guarantees throughput.

I have run this movie before. In 2023 I led a five-person team simulating the digital euro's deposit-shift dynamics under holding limits, and the model returned a 15% potential migration of retail savings out of commercial banks into central bank accounts. The number was not the point. The point was what happened after we presented it in Madrid: a modeled parameter became a policy constraint, and a policy constraint became a balance sheet input within eighteen months. Regulators do not write suggestions. When a duration appears in a rule, someone's treasury function starts hedging it.

The demand side is already aligned. Consumer research puts roughly three-quarters of respondents wanting a human review step on consequential automated decisions. That is not a preference the market can route around. When consumer expectation and statute point the same direction, the architecture gets locked. Fully autonomous agents will not fail because they are unsafe. They will fail because nobody wants to be the first firm that shipped one without a human on the other end.

This Is a Liquidity Problem, Not a Morality Problem

The temptation is to read all of this as a story about vendor ethics. That framing is a dead end, and it wastes the analytical value.

What actually exists is a maturity mismatch. The vendor books revenue on delivery. The customer incurs a floating obligation against that delivery, sized by the agent's error rate, which is unknown at contract signature and drifts as the agent is given more scope. A floating obligation against a fixed revenue line, unmarked, held by the counterparty with the weakest measurement capability — that is the same shape as every liquidity cascade I have modeled.

In 2022 I took apart Terra/Luna and concluded it was not an ideological failure. It was a duration mismatch that nobody marked until the marking became involuntary. Sixty billion dollars of stablecoin value evaporated in roughly 48 hours because a liability with a floating cost met an asset with a fixed promise, and the gap was resolved by the market rather than by the operator. Liquidity doesn't lie. It simply waits for the reconciliation, and then it charges interest for the delay.

Review load has the identical geometry. The liability is real from the first deployment. It is simply unrecognized, and unrecognized liabilities do not shrink — they compound at the rate of adoption. The bear market makes this visible faster, because in a bull market nobody audits their cost base. In a drawdown, every operator runs the same calculation: what did this actually cost me, all in?

Where On-Chain Settlement Changes the Equation

Here is where the crypto conversation stops being adjacent and becomes load-bearing itself.

Payroll cannot price review. It pays for time, not for verified output. A salaried reviewer produces an unverifiable attestation — a Slack message, a checkbox, a comment in a ticket — that cannot be aggregated, transferred, or collateralized. The entire review layer of the AI economy produces value in a format that no accounting system can consume.

Three properties of on-chain settlement fix that, and none of them require a token narrative.

Unit pricing. A signed review record can be paid per attestation rather than per hour. That is the difference between a cost center and a market. Once review is priced per unit, the ratio of review cost to agent-hours becomes observable, and once it is observable it can be optimized, hedged, and disclosed.

Portability and collateral. An attestation that is signed, timestamped, and bound to a specific agent action is an object. Objects can be staked against. A verifier who attests and is later proven wrong can be slashed. That is the mechanism that converts review from an unenforceable promise into a priced instrument with a penalty attached.

Identity separation. I spent the first half of 2025 building exactly this. When AI agents began executing autonomous transactions, the immediate problem was not payments — it was distinguishing a human wallet from an agent wallet, and distinguishing an agent acting under human mandate from an agent acting on its own. We shipped a prototype in three weeks and it raised seed capital from two funds, which tells you the market understood the gap before it understood the product. An attestation from an unverified reviewer is worth nothing. The identity layer is what makes the review layer collateralizable.

The Guardrail Paradox

Before anyone builds a business on staked review, one technical objection has to be answered, and the current marketing language is hiding it.

A guardrail model reviewing an agent built on the same base model family is not an independent check. It is a correlated counterparty. Shared training data, shared tokenizer, shared failure modes. When the reviewed model is confidently wrong in a particular way, the reviewing model is disproportionately likely to be confidently wrong in the same way. The correlation is not incidental; it is structural.

This is the flaw in "an independent model checks the action." Independent in what sense? Different endpoint is not different lineage. I have audited enough contracts to know what correlated failure looks like: seven edge cases in the 0x Protocol v2 codebase in 2018 that no amount of internal review had flagged, because the reviewers shared the same mental model as the authors. Same architecture, same blind spots. The fix was not more review. It was review from a different frame.

A staked reviewer that shares lineage with the reviewed agent is a hedge against the wrong risk. The market will figure this out, and when it does, the premium will accrue to genuinely orthogonal verification — different model families, different modalities, formal methods where the domain permits, and humans where the consequence is legal rather than statistical.

That last category is the durable one. Machine review can be commoditized. Accountable review cannot, because the regulation specifies a person with authority, and a person with authority is a scarce input with a multi-year training cycle.

The Contrarian Read: The Bottleneck Is Not Review Hours

The prevailing thesis — review is an unpriced tax on agent deployment — is directionally correct and structurally wrong, and the error is expensive if you position on it.

The thesis assumes review volume is rigid. It is not. The verifier layer is being automated as fast as the agent layer. Constitutional-style self-critique, critic models trained on execution traces, formal verification for structured outputs — all of it compresses first-pass review. For code generation and structured data tasks, my working estimate is that machine pre-screening absorbs the majority of first-pass review within eighteen months. Not all of it. The majority.

So the cost does not stay where the bearish case puts it. It bifurcates. Cheap machine triage at the bottom, expensive human adjudication at the top. The bottom commoditizes and the price collapses. The top does not commoditize, because the top is defined by legal consequence rather than by task complexity, and legal consequence cannot be delegated to a model that cannot be deposed.

The scarce input, therefore, is not review hours. It is accountable review hours — review performed by an identified party with the authority, the information, and the liability to overturn a machine. That market is far smaller, far less elastic, and far more defensible than the generic "review is expensive" framing suggests.

There is a second error worth naming. The industry tracks inference cost per token, which is falling, and ignores the ratio of review hours to agent-hours, which is the actual adoption ceiling. Every deployment decision inside a regulated firm reduces to that ratio. If one reviewer can supervise three agents, you deploy three. If one reviewer can supervise thirty, you deploy thirty. Nothing about model capability changes that arithmetic. The ratio is the growth curve, and nobody publishes it.

Positioning Through the Drawdown

In a market where survival dominates returns, the question is not which agent protocol has the best narrative. It is which revenue line is exposed to deployment volume and which is exposed to verification events.

Agent-narrative tokens are priced on SaaS logic: high gross margin, near-zero marginal cost, infinite replication. If review load is linear in deployment, that multiple is a spreadsheet error with a market cap attached. This is the single largest mispricing I can identify in the current structure, and it does not require the agents to fail — it only requires them to be supervised.

Compute exposure is real but derivative. Inference demand tracks deployment, and deployment is capped by the review ratio. The capex cycle is not wrong, it is early relative to the constraint that binds it.

Verification and attestation infrastructure is where the asymmetry sits, and for a counterintuitive reason: verification demand is countercyclical to agent reliability. Better agents do not reduce the need for attestation, because the attestation is what makes the output usable in a regulated workflow. Worse agents increase volume. Either way the verification layer gets paid per event while the agent layer gets paid per seat. In a drawdown, event-based revenue with a regulatory tailwind survives; seat-based revenue with an unmarked liability behind it does not.

The same logic explains the exchange layer. Launchpad returns compressed from triple digits to single-digit multiples over successive cycles, which is the market telling you that traffic monetization decays faster than infrastructure monetization. The fee that survives is the one attached to settlement and verification, not the one attached to attention. Every cycle confirms it. Liquidity doesn't wait for the invoice. It moves to where the obligation is priced.

The Signals That Actually Matter

Six months out, watch for three things. Whether Notion or Microsoft discloses review-function usage data — review volume, human intervention rate, escalation frequency — because that single disclosure would validate or destroy the load-bearing thesis. Whether any vendor ships pricing tied to review load rather than seats, which would be the first honest admission that the cost exists. And whether the Harvard paper's full methodology survives scrutiny, because a 49% figure with an undisclosed sample is a narrative input, not an analytical one.

Six to eighteen months out, the Colorado rulemaking is the pivot. If the January 2027 date holds and other states follow, HITL gets locked into American enterprise architecture and compliance review becomes a purchasable service with a statutory demand curve. Watch for the first funded verifier-layer company with a real product rather than a whitepaper. Watch for AI application filings that begin to mention review cost, in risk factors or in MD&A, because that is the moment the liability moves from the customer's payroll onto the vendor's disclosure obligations.

Eighteen to thirty-six months out, one number decides everything: the substitution rate of machine review for human adjudication. Above seventy percent, review is a transitional cost and the bearish case collapses. Below forty percent, review is a permanent structural tax and the application layer's margins never recover. Nobody currently knows which side of that line we land on, and anyone who claims to is selling something.

Takeaway

The reviewer was never free. The reviewer was simply never invoiced, and an uninvoiced liability is indistinguishable from a subsidy until the reconciliation arrives. What is being repriced right now is not the cost of intelligence. It is the cost of trusting it, and that cost has a deadline, a named responsible party, and — for the first time — a settlement rail that can attach a price and a penalty to a signature.

When the first AI application company is forced to disclose customer review cost as a material item, the market will reprice something. The question is which side of the ledger moves first: the application layer that sold infinite scalability, or the settlement layer that finally put a number on the human standing behind the machine. My position is that the application layer reprices within one reporting cycle, and the settlement layer reprices over three. One is a correction. The other is an architecture.