AI’s Value Is in the Decisions Nobody Was Making
Measured in hours saved, AI’s value tends to look small. A bigger share lies in decisions nobody used to have time for.
Most companies work out the return on an AI project the same way: hours saved per person per week, times headcount, times the hourly rate. The number that comes out is usually small and convinces nobody. Over time I’ve come to think the problem is the unit of measurement.
What AI makes cheaper is judgment. Once judgment is cheap, some of the people who make judgment calls today can be replaced. That’s one account. The other is bigger: many decisions that used to be too expensive to make, and so were never made, can now be made. That can’t be measured in hours. It has to be counted in decisions.
Why the hours never add up
The time saved comes in scraps, twenty minutes here and half an hour there. It doesn’t add up to a finished piece of work, let alone one fewer hire.
The bigger problem is that this math treats AI as a helper on the side. The process stays the same, the people stay the same, and the old work just gets done a little faster. What such a project can report in the end is mostly how many people use it and how many requests it handled. What actually changed in the business is hard to say.
Judgment has always been expensive
Every major general-purpose technology takes something only people could supply and turns it into something that can be bought on its own, in almost any quantity. Before the steam engine, power came from people and animals. Before the computer, “computer” was a job: a person who sat in a room doing calculations. When something like that becomes cheap, organizations change shape. This time what’s becoming cheap is judgment, the ability to read through a pile of material and conclude what should be done.
People who judge well have always been scarce, so companies ration them. Many management practices we take for granted have to do with saving judgment.
Budgets are set once a year because every budget round uses up a lot of people’s judgment, and doing it twice a year costs too much. Customers are sorted into tiers A, B and C for the same reason: nobody can look at them one by one. Concentrating decisions in a few hands works in a similar way: front-line judgment can’t be checked case by case, so the front line isn’t asked to judge at all.
There’s nothing wrong with these practices. They just rest on the premise that judgment is expensive.
The decisions left empty
When judgment gets cheaper, the first thing people think of is substitution: let AI do what analysts used to do. That works, but it’s the same math as hours saved. Replace one analyst and the saving is one salary.
The larger gain is elsewhere. A pharma sales rep typically covers more than a hundred doctors. Ideally someone would think about each of them every week. Should I visit this week, what should we discuss, was the question from last time resolved? In practice few people manage it. Reps plan from experience, and managers look at a summary once a month. It matters; it was just never feasible.
Companies have many more decisions left empty like this. With cheap judgment, a decision made once a year can be made every week, and a decision made per region can be made per customer. Filling those gaps is worth far more than replacing an analyst.
Counting decisions
These days I think about the value of an AI project roughly like this:
Value = new decisions made × quality gain per decision − cost of governance
The first term counts decisions nobody made before and someone makes now. It is usually the largest. The second is how much better each decision gets once the system makes it, compared with a gut call or with nobody deciding at all; if it can’t do better, there’s little point handing the decision over. The third is everything spent so that people trust it enough to use it: records, reviews, rollback, clear accountability. Governance that’s too heavy will eat the gains from the first two.
The third term can’t be skipped. When a person’s judgment turns out to be wrong, someone is accountable. When the system’s judgment is wrong, who is? Without an answer, large companies find it hard to really hand decisions over. I set out my own thinking on who should answer for it, and how, in AI Has to Earn Its Decision Rights.
Besides “3 hours saved per person per week,” there’s another question worth asking: once it’s live, which decisions that nobody used to make are now being made?
