Most finance teams point AI at their cheapest work first. Accounts payable, data entry, the high-volume clerical stuff. It feels like the safe place to start: lots of volume, low risk, easy to measure.
A new model from McKinsey suggests that’s often backwards. The best economics for AI aren’t in your cheapest hours. They’re in your most expensive ones.
What McKinsey measured
On October 14, McKinsey published “The AI value curve: The economics of applying AI to existing work.” It’s built on McKinsey’s own model, not a peer-reviewed study, and McKinsey sells AI consulting, so keep that in mind. But the method is laid out clearly, and the method is the useful part.
For 831 occupations, broken out by region and industry, they ran the same calculation:
- Take the fully loaded cost of an hour of that person’s work: wages, benefits, overhead.
- Take what it costs AI to do that same hour of work. That has three parts: the tokens, the human time reviewing the output, and the rework when the AI gets it wrong.
- Subtract, and you get net savings per hour. Multiply by the hours AI can realistically take on, and you get the total value.
If that sounds familiar, it should. It’s a contribution margin chart. Savings per hour is the unit margin, hours is the volume, and margin times volume is the total. Some jobs come out below zero, meaning AI currently costs more than the person.
Added up, McKinsey puts the value at about $2.52 trillion a year globally, with financial services the largest industry at $530 billion. That only covers work people already do today, and it leaves out one-time costs like setup, training and change management.
Why the expensive hours win
When McKinsey sorts by type of job, the best economics aren’t in replacing lower-wage work. They’re in helping high-wage, skilled people. Lawyers, executives, financial managers and software developers sit at the top. Business and financial operations roles sit just below them. Administrative support shows limited value at current wages.
They’re specific that this isn’t the judgment calls or the client relationships. It’s the routine work inside senior jobs: meeting prep, email, pulling together briefings and board materials. Their point is that automatable tasks are not automatable occupations. Nobody is replacing your controller. A chunk of your controller’s week is routine, and that chunk is expensive.
The math is simple. The token cost is about the same no matter whose work the AI is doing. A first draft of board commentary doesn’t cost more in compute because a controller would have written it. But the hour it replaces costs a lot more, so there’s more left over after you pay for the AI and the review.
A worked example
These numbers are made up to show the method.
An AP team of four costs $45 an hour fully loaded. For each hour of work AI takes on, you pay about $3 in tokens, $15 for 20 minutes of review and $7 in rework. That’s $25 of AI cost, so you save $20 an hour. Across 60 hours a week, that’s $1,200.
Now take a controller and three managers at an average of $120 an hour. Their work takes more context, so tokens run about $8. Review is the same 20 minutes, but at their own rate, so $40. Rework is $12. That’s $60 of AI cost against $120, a $60 saving per hour. At six hours a week each of meeting prep, first drafts and report summaries, that’s 24 hours and $1,440 a week.
Twenty-four hours of senior time beat sixty hours of AP. It also holds up better if AI gets more expensive. McKinsey modeled a 30% increase in AI cost, and the global value dropped by about a third, from $2.52 trillion to $1.67 trillion. Run that 30% on the example and AP falls to $750 a week, while the senior case falls to about $1,000.
That doesn’t make AP a bad target. With enough volume it can still win. The point is to run both, because senior people rarely volunteer their own work for automation, and their tasks are harder to see on a process map.
If your work is already offshore
McKinsey’s second finding: AI economics are local, not global. Model prices are roughly the same everywhere, but labor costs aren’t. McKinsey doesn’t mention offshoring, but here’s how I’d apply it. If AP work done offshore costs $20 an hour and your US team still checks the output, that same $25 of AI cost now loses $5 an hour. When someone builds an AI case for offshored work, compare it to what the work costs today, where it’s done today, and check who’s doing the review.
Where to start
McKinsey recommends a task-level audit of your five biggest cost centers. List the roles, the recurring tasks and roughly how many hours each takes. Put three numbers next to each task: the loaded cost of the person’s hour, a rough AI cost that includes review at the reviewer’s rate, and the hours. Multiply and sort. That’s your value curve.
Key takeaways
- Rank AI targets by dollars saved per hour times hours, not by volume or by which project is easiest to start.
- Don’t skip the senior hours. Routine work in expensive roles can be worth as much as clerical automation, or more. Decide up front what the freed hours go to.
- Price review at the reviewer’s rate. On senior work, that’s usually the senior person.
- Compare against what the work costs today, where it’s done today. Offshored work leaves a smaller gap for AI to beat.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: Point AI at Your Most Expensive Hours.
Leave a Reply