Stop Calculating the ROI of AI. Start Questioning Why the Process Still Exists.

Every CFO in America is asking the same question right now.

What is the ROI on AI, and how fast will we see it?

Fair question. Uber blew through its entire 2026 AI budget in four months. The Information broke the story in April and the CTO's own words were "the budget I thought I would need is blown away already." (The Information, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets" theinformation.com) The headlines are full of companies that spent first and asked questions later. So the pendulum has swung hard the other way. Now everyone wants a number. A clean one. Ninety days, three hundred and sixty five days, here's the dollar return.

I get it. I've sat across from CFOs asking exactly that question. But here's the problem. That question only makes sense for one type of AI investment. And most of what's actually valuable doesn't fit neatly into that box.

There are three different kinds of value here, and we keep treating them like one.

The first kind is substitution. You replace a human task with an AI task. A voice agent handles a call instead of an agent. This one is easy. You know your cost per call today. You know what the AI costs. You do the math, and the math either works or it doesn't.

The second kind is coverage. This is where it gets harder. Think about quality assurance. Most contact centers sample two to five percent of calls for QA review, because reviewing one hundred percent with human analysts was never realistic. Too expensive, too slow, too many people you'd need to hire. AI changes that. Now you can review every single interaction. But that's not a cost reduction story. You're not replacing anyone. You're adding coverage that never existed before. And that creates a real problem when someone asks you to prove the ROI.

The third kind is even harder to explain, and it might be the most valuable of the three. Sometimes AI doesn't replace anything or expand anything. It just sees something nobody else could see. A pattern. A broken step in a process that's been broken for years and nobody noticed because nobody was looking at it from that angle. You can't calculate the ROI of that in advance. You don't know what you're going to find until you find it.

I want to walk you through real examples of the second and third kind, because this is where almost every AI ROI conversation falls apart, and almost nobody is talking about why.

Here's what it actually takes to justify the second kind.

A team I worked with wanted to move from sampled QA to one hundred percent AI-driven QA. The case for it seemed obvious internally. Better coverage. Catch more problems. Improve the customer experience. The CFO wanted the dollar number, and fair enough, that's the CFO's job.

So we tried to build it properly.

First we had to establish what the current sample size actually told us, statistically. What was the margin of error on a two percent sample? Then we had to estimate how many service recovery opportunities were being missed entirely, because nobody was looking at ninety eight percent of the calls. Then we had to put a value on what one point of improved CSAT was actually worth. Then tie that CSAT improvement to churn reduction. Then account for the cost of a bad review that shows up on social media or with the Better Business Bureau before you ever hear about it directly. Then roll all of that up into customer lifetime value, in a format a CFO could actually approve.

That is a lot of work. And here's the part that matters most. Most organizations don't have the historical data to do any of it, because human QA was never capable of producing a real baseline. You can't measure the value of full coverage against a baseline that was never built at full coverage. So the business case either becomes a leap of faith, or it becomes a pilot program that limps along without real executive support, or it just dies in committee because nobody can make the spreadsheet work.

None of this shows up in the ROI debate happening in the news right now. (Only 28% of AI projects fully meet ROI expectations according to a Gartner survey of 782 operations leaders, and separately Gartner's CFO research found that leaders must stop mistaking AI deployment for value creation — "they must not mistake activity for impact." gartner.com gartner.com) Nobody is talking about how hard it is to build the case for the investments that don't replace a person.

Here's the thing I keep coming back to. We have gotten so used to good enough that we've forgotten what amazing actually feels like. Every agent is human. They get tired. They have a bad morning. They say the wrong thing on call number forty seven of the day. That's not a flaw in your hiring. That's just what it means to be human. Good service design has always included recovery. Catching the mistake, calling the customer back, making it right before they ever post about it online. The only thing missing was the ability to actually see the mistake in the first place. AI gives you that. The hard part is putting a price tag on something you've never been able to measure before.

Now here's the story that changed how I think about all of this.

Partway through, the AI surfaced something nobody on the team had questioned in years. Three of the sixteen steps existed only because, at some point in the past, a human couldn't track enough information at once to skip them. They were workarounds for a limitation that no longer existed. Once we removed them, cycle time dropped twenty three percent.

Nobody could have written that ROI projection in advance. Nobody knew those three steps were unnecessary until the redesign forced the question. That's the value you can't put in a spreadsheet before you start.

This is the real argument, and almost nobody is making it.

The entire AI ROI conversation right now is built around one question. What is the return on adding AI to what we already do?

That's the wrong question for at least two out of three types of value. The better question is what would we build if we started over, with AI as a first principle, instead of a feature we sprinkle on top of an old process.

Old constraints exist for a reason. They were built around what a human being could see, remember, and do at one time. Sixteen steps because nobody could safely skip three of them without losing track of something important. Two percent QA sampling because one hundred percent was never staffable. Those constraints made sense for decades. They don't automatically make sense anymore. And if you're only asking how AI improves the process you already have, you'll never find the version of the process that wouldn't need to exist at all.

So what do you actually do with this on Monday morning.

Separate your AI investments into the three types before you build a business case for any of them. Calculate the first type the normal way, cost per interaction, clean and fast. For the second type, accept that you're going to need to invest in the baseline data first, and be honest with your CFO that the case will take longer to build because the data has never existed before. For the third type, stop trying to calculate it in advance entirely. Approve it as a strategic bet on discovery, with a defined scope and a hard stop date, so it doesn't turn into an open ended science project with no end in sight.

The companies that figure this out first won't just save money. They'll find process and product opportunities their competitors can't even see yet, because their competitors are still asking AI to make the old process slightly cheaper instead of asking whether the old process should exist at all.

Which of your AI investments are you still trying to force into a ninety day ROI calculation when they're actually something else entirely? Tell me on LinkedIn. I read every comment.