Why you can't prove ROI on AI training (and the number your CFO wants)

You bought the AI tools. You ran the training. A few weeks later leadership asks the obvious question: so what did we actually get for it? And the room goes quiet.

You are not behind. This is the most common conversation in Indian L&D right now. Companies have spent real money putting AI in front of their teams, and almost none can put a number on whether it worked. Three things quietly break the measurement.

• Usage is invisible. You can see licenses bought and logins, not whether people use AI well, rarely, or as expensive autocomplete. A license is an input, not an outcome.
• Trained is not fluent. A workshop attended tells you someone was in the room, not that they changed how they work the following Monday.
• Confidence hides the gap. Heavy users assume they use AI well. The two are not the same, and self-report surveys make it worse, not better.

A well-known field experiment from BCG and Harvard Business School (Dell'Acqua et al., Organization Science) found that on tasks inside AI's capability, consultants using it were markedly faster and produced higher-quality work. On tasks just outside that capability, AI made them worse, and they often could not tell.

The researchers called this the jagged frontier. AI is brilliant on one side of an invisible line and unreliable on the other. This is why headcount using a tool tells you nothing. Your most confident users can be the ones quietly shipping worse work, because they cannot see which side of the line a task falls on. That is what training has to fix, and what you have to measure.

Finance does not want 142 people completed AI training. It wants a rupee figure it can defend in a budget review. That number has a shape.

AI-exposable hours per week, times the cost of those hours, times a realistic productivity lift, adjusted for how ready your team actually is to adopt.

This is a Phillips-style ROI calculation, the same method used to justify any serious L&D spend. It ties a soft-sounding capability, AI fluency, to a hard number: this team spends X hours a week on work AI can accelerate, and closing the fluency gap is worth Y rupees a year. Say that sentence and the AI training conversation stops being a cost debate and becomes an investment decision.

You do not need a year-long study. You need three things.

• Score real fluency, not a quiz. A short conversational diagnostic reveals how a person actually works with AI: what they use it for, where they stop, whether they check the output, whether they can tell a good result from a plausible wrong one.
• Place everyone on a maturity map. Five levels, from not using AI to a multiplier who builds and teaches. Now the whole team is on one picture and the gaps are obvious.
• Price the gap in rupees. Map each level to AI-exposable hours and a realistic lift, and you have the CFO number, per team and per role.

Two weeks, and you move from we think it is going well to here is where we stand and what it is worth to move up.

• A team maturity heatmap: every person, every role, one glance.
• The top three gaps holding the team back.
• The projected rupee ROI of closing them, shown next to what the team costs.
• The specific workshop that closes the biggest gap.
• A re-measure plan so you can show realized ROI, not just projected.

That is the document you hand your CFO. It is the difference between asking for a training budget and presenting a return. Our AI Litmus diagnostic builds it in two weeks, grounded in your team's real workflows and fluency, with every figure labelled by how firm it is.