How to Measure the ROI of an AI Project

"Will it pay for itself?" is the right question to ask of any AI project, and often the hardest to answer. Benefits can be spread across teams, costs hide in cloud bills and staff time, and enthusiasm makes it tempting to count every possible upside. Measuring AI ROI (return on investment) honestly means defining a baseline, capturing the full cost, attributing benefits fairly and being willing to stop if the numbers do not work. This article sets out a practical method, with a worked example using illustrative figures.

The basic AI ROI formula

Return on investment compares what you gain with what you spend:

ROI = (total benefits - total costs) / total costs

Over a defined period, usually one to three years. Alongside ROI, two simpler measures are often more useful for decisions: the payback period (how long until cumulative benefits exceed cumulative costs) and the monthly net benefit once the system is running. The formula is simple. The difficulty is filling it in truthfully.

Step 1: Measure the baseline before you build

You cannot show improvement without knowing the starting point. Before the project begins, record how the process performs today:

• Volume: how many invoices, tickets, orders or documents per month.
• Effort: average staff minutes per item, and who does the work.
• Speed: turnaround time from start to finish.
• Quality: error rate, rework, complaints or write-offs.
• Outcomes: sales conversion, stockouts, customer retention, whichever the project aims to change.

Measure over enough time to capture normal variation, such as several weeks including a month-end. Estimates from memory are usually optimistic.

Step 2: Count all the costs

AI projects have more cost categories than people expect. Use a checklist:

Usage fees deserve special care: estimate them from pilot measurements at realistic volume, including retries and growth, rather than from the provider's headline price per unit.

Step 3: Identify benefits and attribute them fairly

Benefits usually fall into four groups:

• Cost reduction: staff hours saved, fewer errors to correct, lower outsourcing spend.
• Revenue gain: higher conversion, larger orders, fewer lost sales from stockouts.
• Risk reduction: fewer fraud losses, fewer compliance mistakes.
• Capacity: handling growth without hiring in proportion.

Two cautions. First, saved hours only become saved money if the time is redeployed to valuable work or avoids hiring. Otherwise it is capacity, which is real but different. Be explicit about which. Second, attribute carefully. If sales rose during the same quarter as a new recommendation engine, a price change and a marketing campaign, the engine cannot claim all of it. The most reliable way to attribute is a controlled comparison: run the AI for some customers, stores or teams but not others, over the same period, and compare.

A worked example

The following figures are purely illustrative, to show the method; your numbers will differ.

A distributor processes 3,000 supplier invoices a month. The baseline measurement shows about 6 minutes of staff time per invoice, or 300 hours a month. A document-processing pilot shows that after automation, 70 percent of invoices pass straight through after validation, and the remainder take about 2 minutes each to review and correct.

Staff time after:  900 invoices x 2 min           = 30 hours/month
Hours saved:       300 - 30                       = 270 hours/month
Value of time:     270 h x loaded hourly cost of 400 = 108,000/month

Ongoing costs:     AI usage fees                    20,000/month
                   hosting and monitoring            8,000/month
                   maintenance allowance            12,000/month
                   total                            40,000/month

Net monthly benefit:  108,000 - 40,000              = 68,000
One-off build cost:                                 600,000
Payback period:       600,000 / 68,000              = about 9 months

Over two years, that gives total benefits of 2,592,000 against total costs of 1,560,000 (600,000 one-off plus 24 months at 40,000), an ROI of roughly 66 percent. Now test it: what if straight-through rates are only 50 percent in practice, or usage fees double? Running these sensitivity checks shows how fragile the case is. If small changes wipe out the benefit, the project is riskier than it looks. And the saved hours only count as money if the business really redeploys that time or avoids hiring.

Step 4: Track after launch

ROI is not a one-time calculation. Keep measuring the same metrics as the baseline, monthly. Watch for:

• Accuracy drifting down and review effort creeping up.
• Usage fees rising faster than volume.
• Staff working around the system, which hides costs.
• Benefits that were forecast but never appeared.

Benefits that are hard to price

Faster customer responses, better employee experience and more consistent decisions are genuine but difficult to convert into currency. List them separately rather than inventing a figure. Decision-makers can weigh them, but the core case should stand on measurable benefits.

When to stop

Agree a stopping rule before the pilot: if the measured benefit does not reach a specified level by a set date, the project ends or changes direction. Stopping a project that does not pay is a good outcome. It frees budget for one that will. Our software consulting team can help build a business case, and our AI and machine learning development team runs pilots designed to produce the measurements it needs.

Key takeaways

• Measuring AI ROI starts with a real baseline taken before the project.
• Count every cost, especially ongoing usage fees, human review and maintenance.
• Attribute benefits with controlled comparisons, and be clear whether saved time is money or capacity.
• Test the case with sensitivity checks and agree a stopping rule in advance.