Where AI
Actually Creates ROI

Ask most teams where they're spending their AI budget and they'll point at the visible, front-of-house stuff, marketing content, a customer-facing chatbot, the demo that impressed the room. Ask where the measurable returns are, and the answer points somewhere quieter. The gap between those two is where a lot of AI money goes to die.

The value is concentrated, and it's knowable

You don't have to guess where the returns are. McKinsey's analysis of generative AI estimated that roughly three-quarters of the total potential value sits in just four areas: customer operations, marketing and sales, software engineering, and R&D. Not spread evenly across the business, concentrated. That's useful, because it tells you where to look first and where the odds are best.

But 'marketing and sales' being on that list comes with a catch worth understanding.

Where the money goes vs where the return is

The MIT study on enterprise AI found budgets pouring into sales and marketing while the clearest, most measurable returns showed up in unglamorous back-office automation, document processing, reconciliation, internal workflows. The exciting front-office use cases can create value, but they're also crowded, hard to measure, and easy to do badly. The back-office use cases are boring, which is exactly why they're under-invested and why the ROI is sitting there unclaimed.

What makes an ROI-positive use case

The pattern behind the winners is consistent. A use case tends to pay off when:

1. 01The task is repetitive and high-volume. Value comes from doing something a thousand times, not once impressively.
2. 02The output is checkable. You can tell whether it's right, so errors don't quietly compound.
3. 03The workflow can actually change. The tool replaces or accelerates a real step, rather than sitting beside an unchanged process.
4. 04Success is a number. Hours saved, cycle time cut, errors reduced, conversion lifted, something you can put in a before-and-after.

When those line up, ROI is straightforward to see. When they don't, you get a pilot that feels innovative and shows up nowhere in the P&L.

The practical order of operations

Start where the work is repetitive, the output is verifiable, and you can measure the result, which usually means an internal, operational process, not a customer-facing showpiece. Prove the return there, in a number you can point at. Then use that win to fund the harder, more visible bets.

The businesses getting real returns from AI aren't chasing the most exciting use case. They're chasing the one they can measure, and measuring it honestly.