The checker is part of the cost

McKinsey published its AI value curve on October 14 and the coverage caught up with it this week. It is a pricing exercise: 831 occupations, across regions and industries, and one question most AI business cases skip. What does a task cost today, and what does it cost with AI once you include the human who checks the output? The headline is $2.52 trillion a year of existing work where AI already comes out cheaper. The line that matters comes before it: "Many organizations are still measuring AI adoption instead of value. They track users, copilots deployed, and tokens consumed."

Users, copilots, tokens. Those are odometer readings. They tell you the truck moved. They don't tell you whether the load paid.

The cost nobody writes down

McKinsey prices AI as two things: the tokens, and what it calls non-token costs, "including human review and rework." That second line is missing from almost every pitch I've seen. The vendor quotes the per-token price. Nobody quotes the dispatcher who reads the system's draft, catches the wrong appointment time, and fixes it. That person is on your payroll, and in McKinsey's math they sit on the AI's side of the ledger. Where AI plus its checker costs less than the way the work is done today, there is value. Where it doesn't, there isn't, however good the demo looked.

The report borrows a rule from David Tepper at Ascerta that I'd put on the wall: an agent creates value when its probability of success is higher than the ratio of the time it takes to verify its work to the time it takes a person to just do the work. If checking takes half as long as doing, the agent has to be right more than half the time or you are paying twice. Every dispatcher already runs that calculation on a new hire. We just never ran it on the software.

Thirty percent

The effective cost of AI is not fixed, and it is not only falling. Unit prices dropped; usage grew faster. McKinsey cites enterprise large-language-model spending tripling over a 12-month period to the end of 2025. Then it runs one scenario: the effective cost of AI rises 30%, which it calls a moderate move given the price swings already seen. The global value pool falls from $2.52 trillion to $1.67 trillion, a 34% drop. In the United States, $320 billion of a $960 billion pool goes away, the advantage narrows in all 18 industries, and some use cases flip negative: AI costing more than the current way of doing the task.

That is the part an operator should sit with. A business case that only works at today's price is not a business case. It is a bet on the vendor's pricing page. McKinsey's own hedge list reads like an operations manual: hybrid workflows where work can shift between people and AI as the economics move, pricing protections in the contract, keep the human expertise in roles near the threshold, and processes that hold service levels if AI usage has to be cut or suspended. Translation: don't lose the person who knows how to do it by hand until you know what the machine costs in a bad quarter.

Where the curve puts our industry

The strongest economics, per the report, are in augmenting expensive expertise: legal, healthcare practitioners, management, software. Office and administrative support, which is what dispatch, billing and compliance paperwork are, sits lower on the curve. Under US labor costs it is positive. In lower-wage markets it isn't yet. And the hands-on roles the report names, assemblers, inspectors and sorters, HVAC mechanics, sit at the far right where "fewer activities are directly addressable by AI without robotics." Drivers aren't named, but they are the same shape.

So for transportation and logistics the honest reading is this: the back office is near the threshold. Near the threshold is exactly where the checker decides the outcome. Say a system drafts a customer location update and gets it right nineteen times in twenty. That is cheap. The same system at four in five might cost more than the person it was meant to help, because every message now needs a read-through before it goes out. That is not a model problem. It is a measurement problem, and most operations have measured neither number.

My contrarian bit

Everyone will read this report for the $2.5 trillion. The useful number is the one McKinsey says most companies don't have: the cost of the task today, fully loaded, in hours. We had it for inbound email because we counted it before we built anything: three-plus days to react, down to about a minute. Without the first number the second one is a demo.

And McKinsey's last recommendation, plan redeployment before deployment, is the one that will get skipped, because it is the uncomfortable one. If a system frees up four hours a day and nobody has decided what those hours are for, the report says what happens: the productivity "can disappear into lower utilization and unmeasured slack." You paid for the tokens and the checker and got nothing you can point to.

One thing to do this week

Pick one task an intelligent system does for you. Write three numbers. What the task cost per unit before, in fully loaded minutes. What the tokens cost now. How long a person spends checking and fixing the output. Add the last two. If the sum isn't clearly under the first, you don't have a value case. You have an adoption metric.

The vendor sells you the token price. The checker is on your payroll. Until both sit in the same column, you are measuring adoption and calling it value.

Sources: McKinsey (QuantumBlack, AI by McKinsey), "The AI value curve: The economics of applying AI to existing work" (article, October 14, 2026; bottom-up model of 831 occupations by region and industry, fully loaded labor cost vs. token plus human-review cost, 30% cost-sensitivity scenario) · ANI via The Economic Times, "AI could unlock $2.52 trillion in annual global economic value: Report" (October 25, 2026)