The agent case priced a worker who never left

Most agent business cases carry a line for headcount removed. Bain asked 951 companies how much autonomy their production agents actually have. Seven percent run unattended. Thirty-eight percent need a human to approve every action. Another thirty-two percent run inside guardrails and hand exceptions to a person. In roughly seven of ten deployments the human is still in the loop, just not in the middle of the transaction.

Read together, those two facts are not about model maturity. The headcount that left the model never left the cost table. It moved into an approval queue.

How seven percent and fifty-six percent both hold

Start with what each number measures. The Bain sample is 951 global enterprises, and the question was the autonomy level in production, not pilot appetite. Seven percent fully autonomous, thirty-eight percent human approval required, thirty-two percent guardrailed with exceptions going to a person. A small remainder sits lower still, advising but never writing into the system.

Other surveys look incompatible. Wing VC asked 181 senior executives and found fifty-six percent of large enterprises running agents in production. KPMG puts average planned AI spend for the next twelve months at 207 million dollars, close to double the same quarter a year earlier. All three are true because they measure different things: one measures autonomy, one measures whether something shipped, one measures intended spend. Between shipping and autonomy sits a whole layer of people.

Budget writers tend to compress the three into one. A business case needs a savings figure, and the easiest formula is volume times labor cost per unit times an automation ratio. In that formula the automation ratio is coverage, the share of inputs the agent can absorb. The cost denominator that matters is autonomy, the share of inputs no one has to look at again. The two ratios can differ by a factor of three, and the gap lands entirely on the approval queue.

Putting the review minute back into the cost

Take a concrete volume. Ten thousand decisions a month, model calls at twelve cents each, a mid-tier routing price. That parameter is mine, set to show the arithmetic, not a quote. The business case assumes full autonomy: 1,200 dollars of model spend, no labor, twelve cents per decision.

Apply the thirty-eight percent: 3,800 decisions enter the approval queue. At ninety seconds each that is 95 hours. Reviewers are not free. At a fully loaded 45 dollars an hour, the queue costs 4,275 dollars. Total 5,475 dollars, fifty-five cents per decision, 4.6 times the original assumption.

That is one tier. Under the guardrail tier the review becomes sampling, say one in five, and the cost per decision falls to roughly thirty-four cents, still near three times the assumption. The difference is who owns what sampling misses. Full review carries labor only. Sampled review also carries error cost, and error cost is booked when something goes wrong, which means it sits at zero during approval.

The real lever is review speed, not the model price. Cutting model cost from twelve cents to three saves nine cents a decision. Cutting review from ninety seconds to forty-five saves twenty-four cents. Whoever shortens the review is shortening the cost. There are two ways: make the output easier to judge at a glance, or cut the decision small enough that a mistake is survivable. Both are engineering work, usually written into the case as an optimization line with no budget attached.

Ninety percent are raising budgets funded by last round’s savings

Two more numbers from the same survey. Thirty-seven percent of companies targeted cost reductions of eleven to twenty percent. Among those that measured, nearly forty percent landed between zero and ten. Even so, ninety percent are preparing the next round, and the next round buys more autonomy.

Where the money comes from is more revealing. Forty-four percent say the next round is funded by savings from the previous automation program, the largest single answer. Budget growth therefore rests on the assumption that the last round already saved, while four in ten of the companies that measured landed in the zero to ten bucket. Two answers in one questionnaire cannot both be true.

This does not require anyone to be dishonest. Automation savings are never an auditable figure in a finance system. A business unit estimates, finance accepts, and the estimate becomes next year’s baseline. Once it is a baseline it is history, and nobody reopens it. Funding round two from it means using an estimate to guarantee an estimate.

One thing can be done this week. Turn last round’s savings into a table that can be checked: the project, the promised saving, the account line that actually fell in the finance system, and that line’s amount this year. If a project cannot fill four cells, its saving is a valuation, not money. The exercise takes an afternoon, and the next budget usually shrinks by a tier on its own.

Saved hours must clear a thirty percent gate

One more layer gets skipped. Even when review time falls, saved time is not saved money. A two-round interview series covering nineteen enterprises quotes a European pharma services firm plainly: employees saving five to ten percent of their hours usually do not lower project cost, because that fragment is not enough to take on another piece of work. For anyone to be genuinely redeployed, effort on a class of work has to fall by thirty percent or more.

Two reference points sit in the same material. A Canadian consumer goods company requires a 2.5x return at approval, meaning each dollar must show where 2.5 dollars comes from. A European insurer expects five to six times on a claims quality project that lifts manual review from five percent to fifty percent, with recoveries surfacing in the extra volume. In both cases the return lands on a specific account line, not on hours.

So thirty percent is a finance gate rather than a slogan. Saving eight percent of hours never reaches the income statement. Saving thirty-five percent can stop a backfill, and that does. When evaluating a project, ask how much effort has to fall before one person is genuinely freed from this work. Without that number the saving is on paper.

Treat the approval queue as part of the product

This suggests a different way to write the case. Instead of how many people are replaced, state how many decisions one reviewer clears. That ratio is the most honest unit for this wave, because it constrains model quality, interface design and process splitting at the same time.

From the example above, a reviewer at ninety seconds clears forty an hour, and 240 in a six-hour effective day. Of ten thousand decisions, 3,800 need review, which is sixteen person-days, or 0.7 of a full-time role. Put that in the case and several arguments dissolve: halving review volume saves 0.35 of a person, and cutting judgment time from ninety to thirty seconds saves 0.45. Where the engineering money goes is now a comparison between two numbers.

The queue has two side effects worth pricing early. One is backlog. Decisions cluster in business hours, the queue keeps growing after reviewers leave, and the morning shift clears an overnight batch whose context has gone cold, which changes judgment quality. The other is batching. A reviewer sees one item at a time and cannot see how many similar items are behind it, so each decision is judged as a one-off when the batch could have been judged as a class. Neither shows up in a pilot, because pilot volume is too small.

Why no vendor reports review hours

The layer is missed because nobody is paid to write it down. Vendors observe calls, latency and success rates, all collected automatically. Review hours happen inside the buyer’s systems and would have to be reported by the buyer. Vendor pricing runs on calls or seats, so review cost never enters its revenue formula, and reporting it would only shrink the deal.

Buyers have their own inertia. Review time is scattered across many people, a minute here, two minutes there, and it enters no time system and no expense line. Getting the number at approval time requires an observation: sit with a reviewer for an hour and count judgments. Nobody volunteers for that, because it produces nothing that can go into a slide.

So both sides default to its absence. A cost item with no field in either system gets treated as zero at decision time, while it keeps happening in the building.

Our reading

Agent returns are systematically overstated, and the cause is the denominator of the cost model, not model capability. Finance models autonomy, the product delivers coverage, and the difference is a layer of human review with no budget, no headcount line and no target. That layer does not disappear. It moves from labor cost into operating cost, which makes the saving look real on the page.

Two predictions follow, both testable within a year. First, a set of agent projects will be asked for fully loaded cost per decision at renewal, review hours included, and vendors that cannot produce the number will turn a price negotiation into a feature negotiation. Second, approval interface latency will become an explicit procurement criterion, because it converts directly into reviewer throughput, which is money.

Three questions for anyone approving the budget. What share of production decisions needs a human look, and who measured it? How many seconds does one review take, including context switching? Last round’s promised saving, which account line and which number in the finance system? If all three have answers, the numerator and denominator share one definition. If not, twelve cents buys a call, not a decision.

Sources and parameters

Seven percent autonomous, thirty-eight percent approval required, thirty-two percent guardrailed, thirty-seven percent targeting eleven to twenty percent savings, nearly forty percent landing at zero to ten, ninety percent raising budgets, forty-four percent funding the next round from prior savings, forty-one percent naming data access as the top barrier: Bain 2026 Automation and AI Pathfinder Survey, n equals 951, primary. bain.com/insights/your-ai-budget-is-growing-your-returns-arent-heres-why

Fifty-six percent of large enterprises in production with agents, median expectation of fifty percent pilot conversion: Wing VC, State of AI in the Enterprise, n equals 181, primary. wing.vc/content/the-state-of-ai-in-the-enterprise

Average planned AI spend of 207 million dollars over twelve months, roughly double the prior year: KPMG 2026 AI Pulse Survey, Q1, vendor-sponsored survey of stated expectations. kpmg.com 2026 AI Pulse Survey Q1

Five to ten percent of hours saved not converting into cost reduction, thirty percent threshold for redeployment, 2.5x approval hurdle at the Canadian consumer goods company, five to six times projected at the European insurer: two rounds of interviews across nineteen enterprises plus one cross-client consultant, secondary, respondent self-report. fundaai.substack.com/p/deepllm-enterprise-ai-application-2e9

Parameters in the worked example: ten thousand decisions, twelve cents of model cost per decision, thirty-eight percent requiring approval, ninety seconds per review, 45 dollars fully loaded hourly cost. All five are mine, set to show the arithmetic, and none is a vendor quote or an industry average. Every multiplication and ratio above can be recomputed from them.