The Quiet Truth About AI ROI: Measuring Real Returns Without Fooling Yourself

Here is the quiet truth the AI industry would rather you did not dwell on: the overwhelming majority of companies spending money on AI cannot prove it earned them anything. MIT's State of AI in Business 2025 found that 95% of organisations have seen no measurable return on their generative-AI investment. A 2026 Forbes analysis of CEO data found 56% of chief executives reported neither higher revenue nor lower costs from AI in the previous twelve months, and only 12% could point to both. My thesis: AI ROI is not elusive because it is hard to create. It is elusive because most buyers never defined it, never baselined it, and are now too invested to admit it. This article is about not being one of them.

The ROI paradox

Deloitte calls it the paradox of rising investment and elusive returns: spending keeps climbing while measurable payback stays flat. Tens of billions of dollars have gone into enterprise AI, and the returns are concentrated in a tiny minority. That is not a reason to avoid AI. It is a reason to be the minority that measures honestly, because the gap between the 12% who profit and the 56% who see nothing is almost entirely a discipline gap, not a technology gap.

Why ROI stays invisible

When companies cannot show a return, it is rarely because the AI did nothing. It is because of how the decision was made and measured.

• No baseline was recorded. If you never wrote down what the task cost in time and money before AI, you can never prove it got cheaper. You are left with vibes.
• The goal was "adopt AI," not "reduce X." A project aimed at looking modern has no denominator. There is nothing to divide the return by.
• Only the licence was counted. The sticker price is the smallest cost. The real total includes compute (25 to 40 percent higher in our region), integration, staff time, retraining, and the change management to make anyone actually use it.
• Soft benefits were treated as hard cash. "Employees feel more productive" is real but it is not revenue. Confusing the two is how a losing project looks like a winner on a slide.

How to measure AI ROI honestly

Real ROI is not complicated arithmetic. It is honest arithmetic, done before you spend and again after. The formula is old and unforgiving:

ROI = (total value gained minus total cost of ownership) divided by total cost of ownership. If you cannot fill in every term with a number you would defend to your accountant, you do not have an ROI, you have a hope.

Count the full cost of ownership

Count only value you can defend

1. Hard savings: hours removed times loaded wage, or vendor spend eliminated. Countable, defensible, real.
2. Hard revenue: conversion or retention lift you can attribute, ideally proven against a control group that did not get the AI.
3. Soft benefits: speed, morale, capacity. List them, value them conservatively, and never let them carry a business case on their own.

A useful rule for a market where capital is expensive: demand that projected hard value be at least three times the total cost of ownership over 24 months before you commit. The buffer covers the overruns that MIT and S&P consistently document, and there will be overruns.

The bravest answer: don't adopt yet

Somewhere along the way, "we're not doing AI for this yet" became something founders are embarrassed to say. It should be a badge of discipline. For many African businesses right now, the honest ROI calculation returns a negative number, and the correct decision is to wait. Wait until your data is clean enough to be usable. Wait until compute prices for your use case fall. Wait until a cheaper, more mature tool exists, which, given the pace of the field, is often only months away. Choosing not to buy is itself a strategy, and I make the full case for it in this piece on refusing to be pressured into AI before your business is ready.

The cost of waiting is usually a little FOMO. The cost of adopting too early is real money, staff frustration, and a burned budget you cannot re-spend. Those are not symmetrical risks.

Where the returns are actually showing up

It helps to know where the profitable minority found their wins, because the pattern is instructive. The MIT and Deloitte analyses point to returns concentrating not in flashy customer-facing chatbots but in narrow, repetitive, high-volume back-office tasks: document review, reconciliation, fraud flagging, first-line support triage, research and diligence work. The common thread is that these are jobs with a clear before-cost, a measurable output, and enough volume that shaving time off each unit adds up to real money. If you are hunting for your own AI ROI, look there first, at the boring, expensive, repetitive process nobody wants to do, not at the impressive-looking use case that would make a good conference slide. Boring is where the money is, because boring is where you can actually measure it.

The sunk-cost trap

The most expensive words in any AI programme are: "but we've already spent so much on it." That is the sunk-cost fallacy, and AI projects are unusually good at triggering it because the spend is gradual, the hope is renewable, and admitting failure feels like admitting you were fooled. So teams pour good money after bad, funding a losing project because stopping would make the loss visible.

The cure is to decide your kill criteria before you fall in love. Money and time already spent are gone whether you continue or not, so they are irrelevant to the only question that matters: from today, will the remaining spend earn more than it costs?

• Set the kill number up front. "If we are not at X value by month four, we stop." Write it down while you are still objective.
• Review against the baseline, not the hope. Compare to the numbers you recorded before you started, not to how far you feel you have come.
• Separate the decider from the champion. The person who pitched the project should not be the sole judge of whether to continue. They are the most sunk-cost-biased person in the room.
• Treat a clean kill as a win. Recovering the rest of a doomed budget is a good outcome, not a failure. Celebrate it, so your team is not afraid to recommend it next time.

The bottom line

AI ROI is quiet because most of it is imaginary, propped up by missing baselines, uncounted costs, and sunk-cost pride. You do not beat those with a better model. You beat them with three habits: record the before, count the whole cost, and decide your kill criteria in advance. Do that and you will land in the small, unglamorous, profitable minority who can actually prove their AI paid off, while everyone else keeps spending to avoid admitting they cannot. In a market where every shilling of compute costs more than it does elsewhere, that honesty is not just good ethics. It is your edge.