How to Calculate AI Project ROI Before You Build It

To calculate AI project ROI before you build, measure what the work costs today, estimate how much of it the AI system will take over after human review time is netted off, and subtract the full cost of building and running it. Then test the result under conservative, expected and optimistic assumptions. If the conservative case still pays back within a period your organisation accepts, the project is worth building.

What is the ROI of an AI project?

AI project ROI is the net value a system returns over a set period, divided by the total cost of building and running it over the same period.

The formula is simple:

ROI = (value returned minus total cost) divided by total cost

Two related figures matter just as much. The payback period is how many months it takes for cumulative net value to cover the build cost. The net monthly benefit is the monthly value minus the monthly running cost. Decision makers often care more about payback than about a percentage, because it tells them how long their money is at risk.

The hard part is not the arithmetic. It is getting honest inputs before anything exists.

What value should you count?

Count value only where you can trace it to a measured or agreed number. The main sources are:

• Hours returned. Time your team no longer spends on a task, net of the time they spend reviewing the AI output.
• Error cost avoided. Rework, credit notes, duplicate payments, missed deliveries or penalties caused by manual mistakes.
• Speed. Faster quotes, replies or approvals, where you can link speed to revenue kept or won.
• Capacity. Volume growth you can absorb without adding people or overtime.
• Better decisions. Prediction models that reduce overstock, stockouts or waste, valued against what those problems cost today.

Be strict about hours. An hour saved is only worth money if it goes somewhere: less overtime, fewer temporary staff, a hire avoided, or time moved to work with a measurable result. If nobody can say where the hours go, report them separately and leave them out of the headline.

How do you calculate the ROI of an AI project step by step?

1. Pick one task. Define it narrowly, for example “processing incoming supplier invoices”, not “finance automation”.
2. Measure the baseline. Monthly volume, average handling time, error rate and the cost of each error. Use system timestamps or logs where you can, not memory.
3. Set the loaded hourly cost. Salary plus employer costs and overheads, agreed with finance.
4. Estimate the automation rate. The share of cases the system can handle with only a quick human check. Be conservative for messy inputs.
5. Net off review time. Every case the AI handles still takes a person some time to approve. This is the step most estimates skip.
6. List the full costs. Build, integration, data preparation, running costs and maintenance (see below).
7. Calculate payback and ROI over one and three years.
8. Run three scenarios and decide on the conservative one.

A worked example

The numbers below are illustrative arithmetic, not results from any real project.

A finance team processes 2,400 supplier invoices a month at six minutes each. That is 14,400 minutes, or 240 hours a month.

An AI system reads each invoice, extracts the fields, matches it to the purchase order and queues it for approval. Assume it handles 75% of invoices (1,800), and a person spends one minute checking each one: 30 hours. The remaining 600 invoices are exceptions handled manually at six minutes each: 60 hours. The new total is 90 hours, so the system returns 150 hours a month.

At a loaded cost of €40 an hour, that is €6,000 a month, or €72,000 a year.

Now the costs. Assume a build cost of €45,000 and running costs of €1,500 a month (€18,000 a year) for model usage, hosting, monitoring and maintenance.

• Net monthly benefit: €6,000 minus €1,500 is €4,500.
• Payback: €45,000 divided by €4,500 is 10 months.
• Year one ROI: value €72,000, cost €63,000, net €9,000, which is about 14%.
• Three-year ROI: value €216,000, cost €99,000, net €117,000, which is about 118%.

The year one figure looks modest because the whole build cost lands in it. The three-year view and the payback period tell the fuller story.

Which costs do teams forget?

Underestimated costs are the most common reason an AI business case falls apart. Include all of these:

• Integration. Connecting to the ERP, CRM, inbox or document store is often more work than the AI part.
• Data preparation. Cleaning reference data, such as supplier records or product codes, so matching works.
• Review time. The minutes people spend approving output, every day, for as long as the system runs.
• Running costs. Model usage grows with volume. Estimate it per case from a sample of real work.
• Maintenance. Supplier formats change, systems get upgraded and edge cases appear. Budget for ongoing fixes.
• Training and change. Time for staff to learn the new workflow and trust it.
• Risk and compliance review. Data protection checks, and legal review where decisions affect people.
• Ramp-up. Value rarely arrives at full rate on day one. If adoption builds over the first quarter, add two or three months to payback.

How do you deal with uncertainty before you build?

Use scenarios instead of a single number. Change the assumption you are least sure of, usually the automation rate, and keep everything else fixed. Using the example above:

Make the decision on the conservative row. If 18 months is acceptable to your organisation, build it. If it is not, look for ways to reduce the build cost, raise the volume in scope or start with a smaller slice of the task.

You can narrow the range cheaply before committing. Run a few hundred past cases through a quick prototype and measure how many it handles correctly. That single test turns your least certain assumption into a measured one.

How do you value benefits that are hard to measure?

Some benefits are real but hard to price: faster customer replies, more consistent quality, less tedious work for staff. Do not invent numbers for them.

Convert what you can. If slow quotes lose some orders, and you know your average order value, you can estimate the value of a faster response. If manual errors lead to credit notes, last year’s credit notes give you a figure.

Everything else goes in a separate list of qualitative benefits. Decision makers can weigh it, but it should never be what makes the numbers work.

When does an AI project fail the ROI test?

Some patterns predict a weak return before any work starts:

• Low volume. A task done 20 times a month rarely justifies a build, however tedious it is.
• High review time. If checking the output takes nearly as long as doing the task, the saving disappears.
• Unreachable data. When the information lives in a system with no API or export, integration cost dominates.
• Expensive errors. Where one wrong answer costs more than a month of savings, the human review needed may cancel out the gain.
• Nowhere for the hours to go. Savings that cannot be redeployed are not savings.

Saying no to a project with a weak case is a good outcome. It keeps budget for the ones that will pay back. When we run an AI audit, this is the core of the work: quantifying hours saved per task and ranking the results into a roadmap.

How do you check the estimate after launch?

An ROI estimate is a set of claims you can test. Make sure you can:

1. Keep the baseline. Store the original volume, handling time and error figures.
2. Measure the same things after launch, in the same way, so the comparison is fair.
3. Track the automation rate and review time weekly. These two numbers drive the whole calculation.
4. Set stop or continue criteria in advance. For example: if the automation rate is below 50% after two months, pause and fix before expanding.
5. Update the estimate with real figures, and use it to decide what to build next.

Key takeaways

• AI project ROI is net value over total cost, and payback period is often the more useful figure for decision makers.
• Start from a measured baseline of volume, handling time and error cost for one narrowly defined task.
• Net off human review time and include integration, running costs, maintenance and ramp-up, or the estimate will be too optimistic.
• Run conservative, expected and optimistic scenarios and make the decision on the conservative one.
• Keep hard-to-measure benefits in a separate list rather than using them to make the numbers work.
• Test the estimate after launch against the same baseline, with stop criteria agreed in advance.

If you want an honest estimate of what an AI project would return before you commit to building it, talk to us.