Seat licenses aren't ROI. Here's what to measure instead.

Ask most companies how their AI investment is doing and you'll get the same dashboard: seats provisioned, weekly logins, prompts sent, an 'adoption rate' trending up and to the right. None of that is measuring AI ROI. Every one of those numbers can climb while the business gets nothing back. A login is not a result. A prompt is not a deliverable. Adoption rate tells you people touched the tool; it says nothing about whether anything they shipped got faster, better, or cheaper.

The 2025 surveys all found the same gap. McKinsey's State of AI survey put regular AI use at 88% of organizations, yet only 39% reported any enterprise-level earnings (EBIT) impact from it, and roughly 6% could attribute 5% or more of earnings to AI. MIT's GenAI Divide report was blunter: 95% of enterprise gen AI pilots showed no measurable P&L impact. That number went viral, and its critics have a point about method. It counted only direct P&L impact within six months, resting on 52 interviews the report itself called 'directionally accurate.' But notice that both readings land in the same place. Either most AI spend returns nothing, or most companies can't produce evidence either way. To a CFO, those are the same problem.

This is why so many AI investments can't survive a budget review. The CFO asks what the spend returned, and the answer is a usage chart. Usage is an input. ROI is an output, and outputs live in workflows: the proposal that goes out, the close that finishes, the ticket that resolves. If your measurement doesn't name a workflow, it isn't measuring AI ROI. It's measuring enthusiasm.

Even the optimistic numbers prove the point. Wharton's Accountable Acceleration survey of 801 enterprise leaders found 72% now run formal ROI tracking and 74% report positive returns. Look at how, though: the ROI question was answered 'based on internal conversations with colleagues and senior leadership,' and the answers split by rank. 45% of VP-and-above executives called their company's returns significantly positive, against 27% of mid-level managers, the people closest to the actual work. When the reported return depends on the seniority of the person reporting it, you're measuring sentiment. The fix is a different unit of analysis. Stop asking 'is the company adopting AI' and start asking 'what did AI do to this specific piece of work.' Per workflow, before and after, on real runs.

Measuring AI ROI starts smaller than most analytics projects assume. Pick three workflows where output is visible and frequency is high. Sit with the people who run them and time the current version on real work: not the process-doc version, the actual one with the copy-paste steps and the waiting. That's your baseline, and getting it takes a week, no quarter-long study required. Our time-savings calculator gives you a fast first pass on which workflows are worth baselining at all.

Then re-measure monthly. Same workflows, same clock, real runs. Watch all five measures, the first one alone will miss the failure modes that hide in the later ones. A team can post great hours-returned numbers while quality escapes climb, or while every reclaimed hour quietly converts into another standing meeting. If you suspect the second problem, the meeting ROI calculator will tell you exactly where the recovered time went to die.

Now the honest beat. The reason 'we bought licenses' projects can't produce an ROI number isn't that the math is hard. It's that there's nothing to point the math at. No named workflow, no baseline, no before-and-after, just seats and hope. If you can't name the workflow, you can't measure the return, and procurement-first rollouts almost never name one. That's the same root cause behind AI tool sprawl: buying breadth because nobody scoped the work. The companies that can show a number picked specific workflows and trained the people who own them, the sequence in the first 90 days of an AI transformation. McKinsey's data backs the workflow framing: its AI high performers, the roughly 6% attributing real earnings impact to AI, stand out because they fundamentally redesigned workflows, and seat count explains none of it. Measuring AI ROI isn't a layer you add after the rollout. It's what tells you whether you had a rollout at all.