The ROI of artificial intelligence: how to measure AI's impact on your bottom line

Pilots are easy to fund and impossible to justify. A concrete framework for putting a number on AI before and after you build it.

Boards have stopped asking whether to invest in AI. They now ask a harder question: what did the last investment actually return?

Most teams cannot answer, because the pilot was never framed as a financial decision.

Three places AI shows up on the P&L

AI does not have its own line. It moves existing ones:

Cost of delivery. Fewer hours per unit of work. Measurable, unglamorous, and where most real value lands today.

Revenue per rep or per customer. Faster response, better targeting, higher conversion. Harder to attribute, larger upside.

Risk and rework. Errors caught earlier, compliance handled consistently. Invisible until it fails, expensive when it does.

Decide which of the three you are targeting before you build. A project aimed at all three is aimed at none.

The formula, honestly stated

Annual return = (hours saved × loaded hourly cost)
              + (incremental revenue × gross margin)
              − (model + infrastructure cost)
              − (engineering maintenance)

The two terms teams forget are the last two. Inference cost is not fixed — it scales with adoption, which is what success looks like. And a model in production is a system that needs owners, monitoring and periodic re-evaluation.

Budget maintenance at roughly 20–30% of build cost per year. If the return only works at zero maintenance, it does not work.

Measure the baseline or do not start

You cannot claim a 40% reduction in handling time if nobody measured handling time before.

Two weeks of instrumentation ahead of the build is the difference between a result and an anecdote. It is also the cheapest insurance available: sometimes the baseline reveals the problem was somewhere else entirely.

Why most pilots stall

They succeed technically and die operationally. The demo works, and then:

Nobody owns the output in the day-to-day workflow.
The model sits beside the process instead of inside it.
Quality drifts and there is no monitoring to catch it.

An AI feature nobody uses has a negative ROI. It cost money and returned nothing.

Our rule

We do not ship an AI feature without three things agreed upfront: the metric it moves, the baseline value of that metric, and the person accountable for it ninety days later.

That constraint kills a lot of exciting ideas early — which is exactly the point.