Measuring ROI on AI Projects

The biggest mistake businesses make with AI projects is building first and measuring second. By the time you want to know if it’s working, you have no baseline to compare against.

Define the Problem in Measurable Terms Before You Build

“We want to be more efficient” is not a measurable outcome. “We spend 12 hours per week manually processing invoices and want to reduce that to under 2 hours” is.

The difference matters because it tells you:

• What to measure (time spent on invoice processing)
• What the baseline is (12 hours/week)
• What success looks like (under 2 hours/week)
• What the ROI calculation is (10 hours × salary rate × 52 weeks vs. build cost)

Before every AI project we work on, we establish a measurable baseline for the problem the AI is supposed to solve.

The Metrics We Actually Track

Depending on the project, we track some combination of:

Time savings: Hours per week/month recovered from manual work. Easiest to measure, most directly linked to cost.

Error rate reduction: Mistakes per 100 transactions before and after. Relevant for data entry, classification, and processing tasks.

Throughput increase: Volume processed without adding headcount. Relevant for customer support, document processing, lead qualification.

Revenue impact: Harder to attribute directly, but relevant for lead scoring, customer retention, and pricing optimisation.

Pick one or two primary metrics before you build. Track them before, during, and after deployment.

What Good ROI Looks Like

A well-scoped AI project should return 3x the build cost within 12 months from measurable efficiency gains alone. If it can’t clear that bar, it’s probably not the right project.

In practice, the projects with the clearest ROI are the ones that replace specific, time-consuming manual tasks with automated equivalents. The projects with the murkiest ROI are the ones justified with vague claims about “insights” or “competitive advantage.”

The Projects We Push Back On

We push back when a client wants to build an AI system but can’t articulate what manual process it’s replacing or what metric will improve. That’s not scepticism about AI — it’s scepticism about undefined scope.

If we can’t write a sentence of the form “currently, [person] spends [X hours] doing [Y task]; after this project, [metric] will improve to [Z]”, the project isn’t ready to scope.

Build the measurement framework first. Then build the AI.