If the only number you hold AI to is how much it saves, savings is the most it will ever show you. The teams seeing strong returns are the ones that hold it to a growth number as well: 35% of departments that track at least one growth measure report strongly positive returns, against 13% of those that measure AI on efficiency alone, in the latest Wharton and GBK Collective study.
That gap matters now because most leaders are still writing the cost target into their plans. Before you approve next year's AI line, decide which number it has to move.
How should a mid-market company measure AI ROI?
Start with the question that should open every investment case. When I ran digital marketing and analytics for national retailers, every tech pitch started the same way, with a platform and a savings number, and the first question from the table was never about the software. It was "what are we actually trying to fix, and how will we know it worked."
A savings target answers the second half of that question with the easiest number to count, which is hours. The trouble is that hours saved rarely turn up anywhere a board can see them. My line on this is that most AI tools don't take work off the table. They add different work to the same window of time, so the hours move and the team stays just as busy.
Just over half of the leaders in the Wharton study still say AI's main job over the next one to two years is productivity or cost reduction, and only 22% point to growing the business through new revenue, new products or a better customer experience. I understand why, because cost is the line a CFO can verify fastest. It is also the line least likely to change the shape of the company.
Count what AI does that nobody was doing before
The comparison I use is not AI against a person doing the same job. It is AI against no response at all. Think of the inquiry nobody followed up, the quote that went out a day late, the customer question that sat until Monday. Remove those zeros and the result shows up in sales conversion and customer satisfaction, two of the growth measures the study counted. A cost measure never sees it, because nothing was being spent on that work in the first place.
That is also why I keep AI in the same plan as every other growth lever. It competes for the same money and attention as a new hire, a new market or a price change, and it should be held to a number the same way.
Give whoever owns AI a growth number
AI with a named owner and a number to hit does better than AI added to someone's list as one more duty. The same study points that way: among managers and above, 58% at organizations that had designated a new chief AI officer reported strongly positive returns, against 24% where AI was simply added to an existing executive's role.
I do not read that as a case for a new title. In a mid-market company the owner is more likely to be an existing COO or CFO, or an outside operator, and that can work when AI is part of the job they are measured on instead of an extra duty with no number attached. Whoever it is should carry a growth number in the plan, with a date beside it, and be willing to be held to it.
When a savings measure is the right one
Some AI work is a cost play by nature. If the process never touches a customer or a sale, such as matching invoices or cleaning up records, measure the hours and the error rate and leave it there. Just keep that work from becoming the whole AI plan. And if the case for an AI project is that you will need fewer people, that is a project I walk away from, whatever the number says.
Put one growth number next to every AI line
Look at the AI line in next year's plan. If every number beside it is a saving, add one growth measure, such as conversion, repeat orders or revenue from work the team could not have taken on before, and put an owner's name next to it.
If you are working out where AI belongs in your plan, here is how I think about whether your AI plan sits inside your strategy or beside it, and the kind of help I offer mid-market leaders. If you want a second set of eyes on your AI scorecard, I am glad to talk it through.
Source note: Wharton Human-AI Research and GBK Collective surveyed 1,416 U.S. enterprise leaders and entry- and mid-level employees for the study, released October 13, 2026.
John Wyatt · John Wyatt's page · All posts · RSS