In August, two McKinsey senior partners, Hämäläinen and Fuchs (2026), published The decision dividend: How AI creates economic value. It opens with a point I agree with completely: a company knows to the dollar what it spends on labor and materials, and has no idea what it spends making decisions. So I read the rest with a bias in its favor. One disclosure before I start: I spent my consulting years at Bain, so discount my reading of a McKinsey piece as you see fit.
McKinsey's case runs in four steps
Here is the argument in my own words.
First, the cost of a machine-made decision has collapsed. Tasks that took about $40 of human labor in 2023 now cost a fraction of a cent, and by the authors' analysis AI reached cost parity with human labor in about two years. Their comparison is steam power at roughly 55 years, the electric dynamo at 30 and the internet at seven.
Second, cheap intelligence has not yet shown up in most earnings. Companies that bolt copilots and dashboards onto existing processes see under one percent of revenue in EBITDA, while companies that redesign whole workflows around AI report about 20 percent.
Third, that gap cannot come from cutting staff. Sales, general and administrative costs typically run at 5 to 12 percent, the authors note, so even deep cuts there cannot account for a 20 percent uplift. The authors place the value in faster action, better use of the assets a company already owns, and opportunities it would otherwise have missed.
Fourth, the metric. Counting licenses and pilots measures ambition. The article proposes decision throughput instead, defined as "the percent of decisions that are informed, accelerated, or automated by AI," and borrows the factory's logic: judge a plant by what it produces.
The invisible cost is the right place to start
The invisible cost is real. Few companies keep a record of the decisions their leaders make, so the cost of a bad one never reaches a report anyone reads. I built the dice-tax calculator to put a number on it.
Two of the six priorities it gives CEOs deserve a place on the wall. One is to track economic outcomes and ignore activity counts. The other is to treat delay as a cost, illustrated with a $50 million initiative that becomes a $75 million to $100 million problem after four quarters of waiting. A postponed decision usually looks free, because nobody writes down what the waiting costs.
Throughput counts coverage and skips the yield
A metric with no inspection step
A factory is judged on output that passes inspection, and scrap counts against it. Decision throughput has no inspection step. The article concedes the point in one sentence, that "more decisions do not automatically translate into more value," and then leaves the metric without a yield term. A company could move its throughput to 100 percent and decide no better than before.
McKinsey's own research makes this case better than I can. Lovallo and Sibony (2010) looked at 1,048 major business decisions made over five years and asked what predicted good outcomes. They compared the quality of the analysis with the quality of the process: whether the team discussed the major uncertainties, and whether it heard views that contradicted the senior leader. Process mattered more than analysis by a factor of six. An AI-informed decision is, for the most part, a better-analysed decision. Throughput counts the factor that mattered less.
The 20 percent is a ceiling
The footnotes to the article's second exhibit are worth reading. The 20 percent EBITDA figure is self-reported by AI high performers, about 6 percent of the survey's respondents. The randomized trials behind its revenue and customer figures were run on high-growth start-ups, and the exhibit itself says the effect may differ at enterprise scale. The sidebar on costs says the fraction-of-a-cent figures hold for simple, single-step decisions. The direction of the article survives all of that. The 20 percent is what the leaders reached, and a typical company should plan on less.
Leadership decisions fail in a different way
The article's examples are operational, such as a production schedule rerun every hour or demand sensing at a distributor. Those decisions repeat thousands of times, and each one is scored within days. A 40-person company's leadership team makes a few dozen decisions a year that set its direction, such as a senior hire, a new market, a price change or closing a product line. Speed matters there too, but the expensive failure is the wrong call. Nutt (1999) followed hundreds of such decisions in organizations and found that about half of them failed. Making those decisions faster does nothing for that rate unless the way the team decides changes as well.
I should say where this lands for us. Midfire uses AI on every decision a team runs through it, so by the article's measure our customers would score 100 percent. That is exactly why I would not trust the measure on its own.
Count the decisions before you count the AI
The article never says where the denominator comes from. Throughput is a percentage of decisions, and few companies could list the decisions their leadership team made last quarter. That list is the first measurement, and once it exists the rest is a few columns:
- the date the team set for deciding, and the date it actually decided, which turns the article's cost of delay into something you can see
- whether a real alternative was weighed, since Nutt found decisions made with one option in view failed more often
- whether people gave their own read before the leader's view was known, which is the process Lovallo and Sibony measured
- the forecast written down at the time of deciding, with a date to check it, and whether it came true
After a year, the last column tells you whether your faster decisions are also better ones, which is the question throughput cannot answer. Midfire keeps this ledger for a leadership team, though a shared spreadsheet gets you most of the way there.
The inspection is the review date
If I could add one line to the article's six priorities, it would come from the same factory floor the authors borrowed: count the output that passes inspection. For a decision, the inspection is the review date the team set when it decided, and the record of what it expected to see by then.