Ask most executive teams how they're measuring the return on their AI investment, and the honest answer is often "we're not, yet" — followed by a handful of anecdotes about time saved. That's not a measurement framework, and it's a large part of why so many AI programs struggle to secure continued funding past the pilot stage. McKinsey's 2025 Global Survey on the state of AI found that only 39% of organizations can currently attribute any EBIT impact to their AI initiatives, and most of those report an impact of less than 5% of organizational EBIT. Getting to a credible number requires a deliberate measurement approach, not a retrospective explanation.

The vanity metric trap

"Number of pilots launched" and "percentage of employees using an AI tool" are the two most commonly reported AI metrics inside large organizations, and neither tells you anything about value created. They measure activity, not outcomes, and they're popular precisely because they're easy to report favorably. A credible ROI framework has to start further back — at the business outcome the initiative was meant to change in the first place.

Four categories that actually connect to value

  • Efficiency metrics — cycle time reduction, cost per transaction, hours reallocated from manual work to higher-value tasks. These are the easiest to measure and the most commonly reported, but they're only half the picture.
  • Quality and risk metrics — error rate reduction, compliance exceptions avoided, rework eliminated. Often underreported because they're harder to attribute cleanly, but frequently worth more than the efficiency gain itself.
  • Revenue metrics — conversion rate lift, deal velocity, customer retention improvement. Harder to isolate from other variables, which is exactly why they require a defined baseline and control comparison before the initiative launches, not after.
  • Capacity metrics — the strategic work now possible because operational capacity was freed up. This is the hardest category to quantify and the one most often left out of ROI reporting entirely, despite frequently being the largest source of long-term value.

Set the baseline before you start

The single most common measurement failure is not establishing a clear baseline before deployment. Without knowing the pre-AI cycle time, error rate, or cost per transaction with reasonable precision, any post-deployment number is an estimate dressed up as a fact — and it won't survive scrutiny from a CFO or a skeptical board. Baselining should be a required step before a pilot launches, not an afterthought once someone asks for a number.

Leading indicators versus lagging proof

Full financial impact — the EBIT-level number boards ultimately care about — takes time to materialize and is influenced by variables outside any single AI initiative. Leading indicators (adoption rate, task completion accuracy, cycle time in the first 30 days) let you course-correct early and demonstrate progress before the lagging financial metrics catch up. A mature measurement program reports both, clearly labeled as what they are, rather than presenting an early leading indicator as if it were a finished ROI number.

What high performers do differently

Organizations that report meaningful EBIT impact from AI — a relatively small share of the market today — consistently share one habit: they measured before they scaled, not after. They treated the pilot phase as a measurement exercise as much as a technical one, and they used that early data to make a disciplined go/no-go decision on scaling rather than scaling everything that launched.

The number that actually matters

ROI measurement isn't a reporting exercise you run at the end of the year to justify what already happened. Done well, it's the mechanism that tells you, early and honestly, which initiatives deserve more investment and which don't — which is a far more valuable use of the exercise than producing a number for a slide.