Cost and ROI

AI ROI: how to measure value before you spend

Most boards measure ROI after the money is gone. The smart ones measure it before they spend. Here is how, and why the numbers you need are simpler than you think.

By Zakir Hoosen, Director, Fuzzelogic Solutions. Board-level guidance in plain English.

The biggest lie in AI procurement is that you cannot measure return until you have built the thing. That is backwards. You should measure it before you start. If you cannot, you should not start.

I have sat through board meetings where the ROI slide was a projection three years out, built on assumptions stacked on guesses, presented with the confidence of someone reading a weather forecast for July. The board nodded, approved the budget, and twelve months later the same team was explaining why the numbers looked different in practice.

The problem is not that ROI is hard to calculate. The problem is that most teams measure the wrong things, or they measure too late, or they measure only the bits that look good. This article shows you what to measure, when to measure it, and how to spot the numbers that are fiction before they cost you money.

Why most AI ROI is fiction

The starting point is understanding why the numbers in front of you are probably wrong.

Sixty percent. That is not a small failure rate. That is the majority. If your board is looking at an AI proposal with a neat ROI chart, the odds are against that chart being right.

The reason is simple. Most ROI calculations start with the benefit and work backwards. We will save X hours. We will reduce Y costs. We will increase Z revenue. Each number might be technically possible. But technically possible and actually achieved are different things, separated by a gap that most budgets do not cover.

BCG's research found a further problem.

They are spending the money. The return is not there. The issue is not that the technology does not work. It is that the value was never defined clearly enough to measure.

What to measure instead

The boards that get this right measure four things, and they measure them in a specific order.

1. Baseline. What is the current cost, time, and error rate for the process you want to change? If you do not know the starting point, you cannot measure improvement.
2. Constraints. What stops the current process from working better? Is it data, people, systems, or rules? If you do not know the constraint, you will automate the wrong thing.
3. Pilot results. What happened in the first twelve to sixteen weeks? Measured against the baseline, not against the projection.
4. Scale economics. What does it cost to go from pilot to full rollout, and does the per-unit cost go down or up?

The baseline is the one most teams skip. It feels like busywork. It is the most important number you will collect. Without it, you are comparing a future state against a guess, and a guess against a guess is not measurement. It is fiction.

The test that saves you money

Before you approve any AI spend, run this test. Ask the team proposing the project to answer three questions.

1. What specific process changes, and by how much?
2. What is the cost of the process change, not just the technology?
3. What happens if we do nothing for twelve months?

If the team cannot answer the first question with numbers, the project is not ready. If they can answer it but have not costed the process change, the budget is wrong. If they cannot answer the third question, the project might not be necessary at all.

That third question is the one that saves the most money. Not every problem needs AI. Some need better data entry. Some need a different workflow. Some need a conversation between two departments that have not spoken in years. AI is expensive for problems that do not need AI.

Two thirds of those transformations did not deliver. If you had asked the "what happens if we do nothing" question before those programmes started, a good number of them would have been stopped or redirected. That is money saved.

The OpenAI lesson on measuring outcomes

OpenAI has recently pushed a idea that is worth paying attention to: cost per accepted outcome. Not cost per query. Not cost per token. Cost per outcome the business actually accepts as useful.

That shift matters. Measuring cost per query tells you nothing about value. A thousand queries that produce nothing useful are worse than fifty that solve a real problem. Measuring cost per accepted outcome forces the project to prove its worth in terms the business understands.

You can apply the same principle internally. For any AI project, ask: what is the cost per decision the business accepts? If the system makes a hundred decisions and the business accepts sixty, the acceptance rate is sixty percent and the true cost per useful decision is much higher than the headline number.

This is not complicated. It is just honest. And honesty before you spend is cheaper than honesty after.

Where boards go wrong on ROI

The pattern I see most often is the ROI that lives in a slide deck and never gets revisited. The project is approved based on a projection. Six months in, the numbers are not tracking. Nobody updates the board. Twelve months in, the gap is too large to ignore, and someone has to explain why.

The antidote is simple. Build the review into the approval.

1. Set a fixed review point at twelve weeks. Compare pilot results against the baseline. If the numbers do not support continuing, stop. If they do, agree the next milestone and the next review. Keep doing this until the project either proves its value or earns the right to be stopped.

Fuzzelogic works to twelve to sixteen weeks to a working MVP. That is long enough to prove whether something works in your environment with your data. It is short enough that if it does not work, you have not wasted a year and a budget. You have spent a defined amount, learned what works and what does not, and you own the verdict either way.

The honest version

Measuring AI ROI before you spend is not hard. It requires a baseline, a clear definition of what success looks like, and the discipline to stop when the numbers do not add up. What makes it hard is that most teams do not want to measure before they build, because measurement might tell them something they do not want to hear.

That is exactly why you should measure. The cost of finding out early is a fraction of the cost of finding out late.

Start with the assessment. Two to four weeks, fixed price, and you own the verdict and the roadmap whether or not we build any of it. When you are ready to talk AI, call Fuzzelogic Solutions and ask for Zak. www.FuzzelogicSolutions.com | info@FuzzelogicSolutions.com | +44 (0)1624 618950

Want a practical framework for working out what to measure? Download our free guide, The Boring AI Fix. For more on what AI actually costs, read How much does AI implementation cost, then The 10-20-70 rule. The full library is on our index. Our site explains how Fuzzelogic approaches AI for business. You can reach Zak directly via our contact page.