The ROI tree: how to measure AI ROI immediately

There is a question I hear from almost every leadership team weighing an AI investment, and it usually arrives in nearly the same words. "We track hours, so we can see when a job goes faster. But when we experiment with AI, it sometimes takes just as long as doing it ourselves. How are we supposed to quantify any of this?"

It is a fair question, and the frustration behind it is real. Most teams try to measure AI at the exploration stage, when people are poking at a chat window, going down rabbit holes, and producing work that took the same time as before with a few extra detours. Measured there, AI looks like a wash. The mistake is not the tool. The mistake is where the stopwatch is pointed.

Exploration is learning, and learning is worth the time, but it is not where the return lives. The return lives in workflows: the specific, repeatable processes your team runs every week that you codify into the system once and then run over and over. When you measure at the workflow level instead of the exploration level, the ROI question stops being philosophical and starts being arithmetic.

Two layers, in order

Before we get anywhere near revenue, I ask teams to picture the return as a tree with two layers. The first layer holds the immediate, clear returns, and there are only two of them: velocity and capabilities. The second layer holds the business metrics everyone actually cares about. You reach the second layer through the first, deliberately, not by hoping it shows up on its own.

Layer one, first return: velocity

Velocity is the quantifiable half of the first layer. It is everything your team already does today, done meaningfully faster. The routine reporting, the manual data pulls, the multi-handoff production work, the late-night formatting. When a process that took ten hours of hands-on effort runs in fifteen minutes with a human review at the end, that is not a rounding error. That is a 2x, 4x, or 8x change in throughput, and it shows up as time saved, hours recaptured, and cost pulled back into the business.

The measurement trick is that a codified workflow carries its own baseline. When you teach the system your process, you are documenting the manual steps as part of the work. Once dozens of workflows are codified, the system itself can report the delta: here is what this took manually, here is what it takes now, here is the multiplier. At a prior employer, a large-scale fintech, I helped teams stack more than a hundred codified workflows this way, and velocity reporting became nearly automatic. No stopwatch, no time-tracking theater. The baseline was baked into the build.

Layer one, second return: capabilities

Capabilities are the qualifiable half, and they are easy to undercount precisely because they have no baseline. This is work you simply could not do before, so there is no "before" to compare against. You cannot measure the velocity of something that used to be impossible.

But you can absolutely see it. It shows up as new services and offerings you could not have staffed. Higher and more consistent quality, because the system applies your best practice every time instead of whenever someone has the bandwidth. Expanded scope on existing work. And, over time, work your competitors cannot match at your price or your speed. For services businesses especially, this is the layer that changes what you sell, not just what it costs you to deliver.

Layer two: the metrics that follow

Revenue, retention, engagement, client acquisition, share of market, speed to market. These are the metrics leadership meetings are made of, and AI can move every one of them. But nobody moves them by pointing a model at "revenue." You move them by deciding, in advance, what you will do with the velocity and capabilities you gain.

This is where the reinvestment decision matters more than any technical choice. When velocity jumps, there are two paths. The first is to do the same amount of work with fewer people. It is the easy way out, and I have watched large enterprises take it and quietly cap their own upside. The second is to reinvest the recovered time into more value: more thought leadership for clients, more experiments, new offerings built on the capabilities layer. Do more with more. Teams that take the second path are the ones whose layer-two metrics actually move, because the time saved becomes the fuel for growth rather than a line item in a cost review.

Feel it in the first week

The practical consequence of all this: pick your first workflow so the ROI is felt immediately, not promised for next year. The best first candidate is high-frequency, painful, and simple to codify. For most teams that is reporting. Export the data you already export, hand the system prior reports as examples of a good outcome, give it your brand guide and what the metrics mean, and codify the steps. A weekly client report that consumed most of a day becomes ten minutes of generation and five minutes of human review, and the person reviewing it is the same expert who used to build it by hand.

One workflow like that does three things at once. It produces a velocity number you can put in front of anyone. It teaches the team how codifying a process works, so the next workflow comes faster. And it earns the confidence to go after the capabilities layer, where the compounding really starts.

So when someone asks how to measure the ROI of AI transformation, my answer is: immediately, if you measure the first layer first. Velocity you can count. Capabilities you can name. And the business metrics you care about most stop being a leap of faith, because you can trace exactly which workflows and which new capabilities are feeding them.