How to Estimate ROI for an AI Project

Every AI business case has the same two failure modes, and they usually appear together. The cost side counts the model and nothing else. The benefit side counts a percentage improvement nobody can trace to a workflow.

The result is a number that survives the steering committee and dies in delivery.

Here is a method that holds up.

Cost the whole system, not the model

Inference is rarely the largest line item. These are, roughly in order of how often they are forgotten:

• Integration. Connecting to the systems the workflow actually touches. This is usually the biggest single cost and the least estimated.
• Human review. If a person checks the output before it is used, that person's time is a cost of the system, not a cost of the old process.
• Evaluation and monitoring. Building the ability to know whether outputs are correct, and keeping it running.
• Inference. Real numbers, but usually smaller than teams expect, unless the workflow is high volume.
• Data preparation. Extraction, cleansing, and the unglamorous work of agreeing what fields mean.
• Change and training. People have to learn a new process, and some of them will resist it.
• Maintenance. Prompts, models, and integrations all drift. Someone has to own that.

A business case that lists only the fourth item is not a business case. It is a licence fee.

Count benefits you can trace to a workflow

Defensible benefits look like this:

• Hours removed from a named task performed by a named team, at a stated volume.
• Exceptions handled without escalation, measured against the current escalation rate.
• Cycle time reduced on a step where cycle time is measured today.
• Volume absorbed without adding headcount, at a growth rate you can point to.

Notice what each of these has in common: there is a baseline, there is a measurement, and the benefit is in the workflow rather than in the model.

Indefensible benefits look like this: "improved decision quality", "better customer experience", "increased employee satisfaction". These may be true. They are also unfalsifiable, and putting them in a base case is how a business case becomes fiction.

Put them in the qualitative column. Mention them. Do not sum them.

Build three cases, not one

A single-point estimate invites an argument about whether the number is right. Three cases invite a decision.

• Base. Conservative costs, conservative benefits, no assumption that adoption is immediate.
• Upside. Things go well: integration is easier than expected, adoption is fast.
• Downside. Integration is harder than expected, review effort is higher than expected, and the scope narrows to survive.

The downside case is the useful one. If the project only works in the upside case, you have learned something important before spending anything.

Price the human in the loop honestly

This is where most estimates go wrong. If a reviewer spends two minutes on every output and the volume is 5,000 a month, that is roughly 167 hours a month of new work created by the system. At a loaded rate, that number is often larger than every technical cost combined.

Sometimes that review is still worth it, because the alternative is worse. But it has to be in the model, priced, and compared against the labour it replaces.

The number that matters: cost per unit of work

Total cost divided by volume of work handled gives you a number you can compare against the current cost of doing that work. It also gives you something a percentage never does: a line that should fall over time. If cost per unit of work is not falling as the system matures, something is wrong.

What to do when the estimate says no

Sometimes it does. That is a good outcome, delivered early and cheaply.

When the numbers do not work, three things are usually worth trying before abandoning the use case: narrow the scope to the highest-volume slice, remove the human review step where the risk genuinely allows it, or find a cheaper model and spend the savings on retrieval quality. If none of those move the number, the honest answer is that this workflow is not an AI project yet.