The Missing Ledger: Agentic AI Is Billing You for Its Own Doubt
Salesforce has booked 7 billion agentic work units. MIT's NANDA program measured 95% of generative AI pilots failing to move a single financial needle. Both numbers are true. Only one of them is a metric.
If a protocol settled 7 billion transactions and 95% of them never reached a useful final state, no auditor would call that growth. They would file an incident report, isolate the failing opcode, and halt withdrawals. In enterprise software, the same evidence ships as a $1.5 billion annual run rate and a keynote slide. The gap between the counter and the ledger is the story, and I have spent the last decade auditing the on-chain version of it — from a Golem task-distribution race condition in 2017 to the Compound oracle failure in 2021. The failure signature is always identical: activity gets measured, value gets assumed, and nobody reconciles the two until the treasury is empty. The current agentic AI cycle is running that same script, just with better staging.
Salesforce's Agentforce is the reference case. Its momentum claim rests on two figures: 7 billion Agentic Work Units consumed, and $1.5 billion in annual recurring revenue. The first is proprietary — a custom unit defined by the vendor, incomparable across platforms, unreproducible on a rival stack. The second is financial, and it is the one that enters a valuation model.
Between the two sits a definitional void. What is an AWU worth to the buyer? The vendor does not say, and the vendor does not have to. Over three years the enterprise AI narrative migrated from "look what the model can do" to "look how much of it we are doing." Volume replaced value as the headline. That substitution is not accidental. It is the same move I dissected in ICO-era whitepapers — the ones that counted transactions per second while quietly ignoring settlement finality, because finality is harder to put on a slide. The marketing team learned nothing from 2017. It simply changed the vocabulary.
The buy side is not confused, and this is where the data turns cold. The Futurum Group reports that enterprise decision-makers have shifted from productivity abstractions to direct financial impact. McKinsey found 93% of enterprises overspending on AI against their own budgets. S&P and MIT found that 80% of applications now embed AI, but only 31% run agents in production. Gartner, never shy about pre-announcing disappointment, projects that 40% of agentic AI projects will be cancelled outright.
Read those four sources together and a single structure emerges. Structure reveals what emotion conceals. The structure here is plain: the seller measures activity, the buyer measures outcome, and the two definitions are drifting apart faster than any dashboard reconciles them. This is not primarily a marketing problem. It is an architecture problem wearing a marketing costume, and the costume is starting to slip.
There is a fifth party in the room that no one audits: the interpretation layer. Gartner issues the cancellation forecast. McKinsey diagnoses the overspend. Futurum tracks the ROI pivot. These firms are not neutral observers of the metric confusion — they are its beneficiaries. The messier the measurement, the larger the demand for firms that sell "meaning." When an analyst house calls for standardization, it is also calling for a seat at the table that writes the standard. That conflict is disclosed nowhere, and it should temper how much weight any single one of these projections carries.
Start with the most important number in the dataset, because almost everyone reads past it: 60% of agent spend goes to iterative response optimization — the check, correct, and improve loop. That is McKinsey's figure, and it is an architecture confession dressed as a cost breakdown.
An agent is not a single inference. It is a composite system: multiple reasoning passes, external tool calls, and a verification layer that grades and rewrites its own output before anything is released. The frameworks have names — ReAct, Reflexion, verifier-generator loops — but the economics are simpler than the acronyms. The cost of an agent's answer is not the cost of producing it. The cost is the cost of doubting it, over and over, until the loop finally terminates. Sixty percent of the bill is the machine arguing with itself in a language it invented, and the customer receives an invoice in a unit they cannot define.
I watched this pattern in code before it had a marketing term. In 2025, I audited the first wave of autonomous AI-agent contracts on Ethereum. The finding was structural, not incidental: non-deterministic model outputs were being piped into deterministic state transitions. The agent proposed an action, the contract executed it, and a verification step — a human or a script — was the only guardrail between a hallucination and a settlement. That verification step was also the most expensive component of the pipeline. It still is. The industry did not solve the problem; it renamed it "test-time compute" and started billing for it.
This produces a measurement problem that no KPI layer can patch. Agent output is a stochastic process. The same agent, given the same class of input, produces output with wide quality variance. A deterministic counter — work units consumed, records processed, tickets closed — cannot describe a random variable. It can only describe how many times the variable was sampled. You are not measuring quality. You are measuring volume, and volume is precisely what a broken loop maximizes. An agent that fails three times before succeeding will register four units of activity and one unit of value. The counter sees four.
Connect the cost to the reliability cliff and the arithmetic turns ugly. Eighty percent of applications embed AI. Thirty-one percent run agents in production. The missing 49 points is not an adoption gap — it is a reliability-engineering gap, and it has a precise technical shape. In a controlled pilot, tool calls succeed, context stays clean, and multi-step error does not compound. In production, tool endpoints degrade, context windows fill with noise, and a two-percent per-step error rate becomes a 19% failure rate across ten steps. This is distribution shift, and it is unforgiving. The pilots that worked were never proof the agent was reliable. They were proof that the environment was controlled.
Truth is found in the hash, not the headline. A production agent that fails one in five multi-step tasks is not a productivity tool. It is a liability with a latency budget, and its cost is being billed to the buyer as if the failures never occurred — because the metric used to bill them structurally cannot see failure at all.
Now follow the budget, because money always tells the truth even when metrics do not. More than half of generative AI spend flows to sales and marketing use cases. The most measurable return, by the industry's own evidence, consistently appears in back-office automation: document handling, invoicing, reconciliation, compliance checks. The capital is chasing the demo and abandoning the ledger. It is flowing to the scene with the best narrative and away from the scene with the cleanest accounting. In any market I have audited, that is the textbook definition of a mispriced asset.
Why does the activity metric survive when the outcome data contradicts it? Because the proprietary unit is a moat. An AWU cannot be compared to a competitor's unit, cannot be audited by a customer, and cannot be reproduced on another platform. It functions as a migration barrier dressed as a growth metric. When a vendor defines the unit of success, the vendor defines the terms of the argument, and the customer is left holding a number they cannot verify. Gartner's own analyst publicly conceded that these projects are still hype-driven. Futurum's analyst insists the market has matured. Both can be right simultaneously, because neither is measuring the same object.
I have seen this exact split in DeFi oracles, and it is worth naming because the pattern repeats without mercy. Chainlink was celebrated for decentralizing price feeds while operating them through a permissioned node set. The decentralization claim and the operational reality lived in separate documents, and the gap between them was the attack surface. Feed latency was the Achilles' heel of DeFi lending. Correction-loop cost is the same heel for agentic software. Both are invisible in the headline and decisive at settlement.
The transfer of that cost is the buried headline. Sixty percent of spend funding self-correction is not a rounding error. It is a structural transfer of the agent's inefficiency from the vendor to the buyer. When the vendor charges per token or per activity unit, the customer pays for every failed attempt the agent makes. The vendor's margin is insulated from the agent's accuracy. The buyer's cost is not. This is why sophisticated enterprises are demanding cost-per-outcome pricing. It is not a procurement fashion. It is an attempt to stop paying for the machine's uncertainty, and to push the correction loop back onto the party that designed it.
Now the second-order effect almost nobody has priced: compute. An agent's token consumption is not conversation-level; it is task-level. A single agent task may burn five to fifty times the tokens of a single chat turn, and the multiple grows with task complexity because the loop grows with it. The same daily-active-user count that loads a chatbot lightly can load an agent deployment an order of magnitude harder. That is an under-modeled increment in inference demand — and, simultaneously, an under-modeled ceiling. If correction-loop cost does not fall, agentic AI's addressable market is capped by the price of inference, not by the cleverness of the model. The economics will not negotiate with the roadmap.
The levers that matter are therefore not model size. They are loop count and cache reuse. A better verifier that terminates a loop one step earlier, a smaller model doing the grading, KV-cache reuse across repeated contexts — each of these moves cost-per-outcome more than another billion parameters ever will. The vendors who win the next phase will not be the ones with the smartest model. They will be the ones whose agents doubt themselves the least, and who can prove it.
Two questions remain unmeasured, and their absence is itself a finding. First: are these activity figures ever independently verified? A self-reported unit from a single vendor, uncross-checked by any third party, is not evidence — it is a press release with a decimal point. Second: when an agent executes a wrong back-office action, who carries the liability? I have yet to see a single agent metric that tracks error rate, over-permissioned behavior, or data leakage. Activity is counted meticulously. Failure is not counted at all. A system that counts its successes and ignores its failures is not a measurement system. It is a brochure.
Here is what the bulls got right, and it is not nothing. The back-office ROI is real. Every credible dataset points the same direction: document processing, reconciliation, and compliance automation deliver measurable financial return while the flashy front-office agents deliver anecdotes. That is not a failure of agentic AI; it is a failure of where the budget went. An operator who reallocates spend toward the boring, verifiable, high-frequency back-office tasks is not being conservative. That operator is early to the only segment with a proven ledger, and is quietly capturing returns that the front-office crowd will not notice until the budget is already spent.
The second concession: cost-per-outcome is the correct destination, even though it will be gamed. The demand for outcome-based pricing is maturity, not cynicism. It forces the vendor to absorb the cost of the correction loop, which is exactly the incentive needed to make the loop cheaper. When the seller's margin depends on the agent being right the first time, verifier research stops being a cost center and becomes the product. The obvious counter-risk is that whoever defines "outcome" will eventually manipulate it, which is why the definition needs a third-party audit, not a vendor's self-assessment. But the direction is correct.
And the third: Gartner's 40% cancellation figure may be the most bullish number in the entire article. Half the projects on the books have no business existing. Clearing them removes the noise currently drowning the signal and forces the survivors to publish unit economics instead of activity counts. A bear market is not the enemy of the technology. It is the audit the technology was always going to get. The protocols that survived 2022 did not survive because sentiment improved. They survived because the math worked when sentiment did not.
So the question is not whether agentic AI works. It is whether anyone can prove it before the budget is exhausted and the board asks the only question that ever matters: what did we buy? You cannot audit what you refuse to measure, and the industry is still measuring the wrong thing — in part on purpose, because the wrong metric flatters the invoice. Watch two variables over the next twelve to eighteen months. First, who defines cost-per-outcome, and whether that definition gets independently audited. Second, how fast the correction-loop cost curve actually falls. If neither moves, then "activity is a liability" stops being an analysis and becomes a line item in someone's next earnings call. The wallet does not lie. Only the dashboard does.