Why a single training ROI number is usually fiction

The slide said ROI: 312%. The L&D manager had done what the brief asked. Then the CFO asked two questions. What is the 312 divided by, and is the extra revenue counted at invoice value or at margin? The meeting moved to the next slide, and the number was never quoted again.

Nothing dishonest happened there. Take the benefit, subtract the cost, divide by the cost, multiply by 100, and you have followed the standard method correctly. The formula is not the problem.

The problem is that three of its inputs are not measurements. They are judgements made late, by whoever is writing the deck, and each one can swing the answer by more than the answer is worth.

The formula is sound. Its three inputs are guesses

Written out, the calculation looks like arithmetic:

ROI (%) = (net program benefits ÷ program costs) × 100

Every ROI guide uses it, including the Phillips methodology that most Indian L&D teams are trained on. Phillips' own description of the process treats isolation, conversion and costing as three separate steps, each with its own standards. That is the tell. If the inputs were facts, you would not need a standards document for each one.

Three judgements sit inside the arithmetic:

1. How much of the observed change the programme caused. A sales metric rose. Training was one of six things that happened that quarter.
2. What one unit of that change is worth. An extra order is worth revenue, or gross margin, or contribution after a cost allocation. Three different numbers, same order.
3. What counted as a cost. The invoice is the smallest part, and it is the only part most reports include.

Pick the generous option on all three and you get a number that survives a meeting. Pick the conservative option and the same programme looks mediocre. Both calculations can be arithmetically perfect.

Build the cost line first, because it is the only part you can know

Cost is the one input you can actually close. Phillips' list of cost components includes design and development, materials, the facilitator's preparation as well as delivery time, facilities, travel and lodging, the salaries and benefits of participants while they are away from work, and allocated administrative overhead.

That list looks bureaucratic until you apply it. Participant time is usually the largest single line, and almost nobody counts it, because it does not appear on any invoice.

An illustrative two-day sales programme for 80 people:

The per-head cost is ₹12,390, not ₹4,500. The vendor fee is 36% of the true cost.

This is where most ROI arguments are actually won or lost, and it has nothing to do with statistics. A programme quoted at ₹4,500 per head needs a smaller benefit to pay back than one that really costs ₹12,390. If your denominator is 2.7 times too small, your ROI is 2.7 times too generous, before you touch the numerator.

Two rules keep the cost line honest. Include the hours nobody invoices, and use the loaded salary, not the gross. A ₹6 lakh annual package at 2,000 productive hours is about ₹300 per hour once benefits and overhead are added, and that is the number to multiply by hours away from work.

Pick the isolation method before you book the trainer

Isolation is the step that decides how much of the change belongs to training. Phillips is blunt that it is often overlooked, and his standards require at least one method to be used.

There are four methods in ordinary use, and they do not produce the same answer.

The methodological literature is clear that most real programmes sit in joint attribution, not sole attribution: the intervention contributes alongside other changes, and asking "how much did training cause" invites a number that cannot exist in the form requested. That is why the honest output is a contribution estimate with the method named, not a causal verdict.

Name the method in the report. "Benefit estimated at 40% attribution, from manager estimates across six reporting managers, discounted by a quarter for optimism" is a sentence a CFO can interrogate. "ROI: 312%" is not.

One unit of improvement needs a value you can defend

Converting the metric to money is where the quiet inflation happens. Revenue is not benefit. An extra order adds margin, not turnover, and in most Indian B2B businesses the margin on incremental volume is well under half the invoice value.

Three habits keep this step clean:

• Use contribution or gross margin, never revenue. If the unit value is contested, use the lower figure and say so.
• Take the unit value from a system, not from memory. Cost per rework hour, cost per failed collection, margin per order, cost per exit. If no system holds it, use the most conservative internal expert estimate and label it as an estimate.
• Count the first year only for a short programme, which is Phillips' own standard. A three-year benefit projection on a two-day workshop is a forecast wearing a lab coat.

Then round down. Rounding down costs you a few points of ROI and buys you the room to be believed.

The same programme, three honest ROIs

Back to the 80-person programme at ₹9,91,200. Suppose the measured effect is a two percentage point lift in conversion on 3,000 qualified enquiries a month, at ₹7,500 gross margin per converted order. That is 60 extra orders a month, ₹4,50,000 a month, ₹27,00,000 across six months.

Now vary only the attribution share, holding every other number fixed.

Same programme. Same measured lift. One number says the programme was excellent, another says it lost money. The only thing that changed is a judgement about causation, and nobody in the review will ask about it if the slide just says 172%.

This is the whole case against the single number. It is not that ROI is impossible. It is that a point estimate hides the one input that carries most of the uncertainty.

Report a band, and say what moves it

The fix is small and it takes ten minutes. Replace the point estimate with three lines.

• Cost: fully loaded, with participant time on its own line. This is a hard number and it should be stated to the rupee.
• Benefit range: the same metric at a low, central and high attribution share, with the isolation method named. Show the arithmetic in one line so it can be checked.
• Decision rule: what result would change the decision. "We extend to the second region if the lift holds above 1.5 points at 30% attribution."

A band is harder to misquote and easier to defend. It also changes the conversation. Instead of arguing whether the number is right, the room argues about the attribution share, which is a real question with a real answer, and one you can improve next cycle with a comparison group or a longer trend line.

The one number worth defending is break-even

If you want a single figure on the slide, make it the break-even threshold, because it inverts the problem. Instead of estimating the benefit and hoping the attribution is accepted, you ask what attribution the programme needs in order to pay for itself. That is arithmetic on numbers you already have.

Break-even share = fully loaded cost ÷ benefit at 100% attribution

For the programme above: ₹9,91,200 ÷ ₹27,00,000 = 36.7%. So the line becomes: "This pays back inside six months if training caused at least 37% of the observed lift. Our best estimate is 40% to 50%, from manager estimates. The estimate is the risk, not the arithmetic."

That sentence is checkable, falsifiable and short. It also survives the follow-up question, because the follow-up is about the 37%, and 37% has a visible derivation.

The same inversion tells you what the metric has to do. At 30% attribution, each percentage point of conversion lift is worth ₹4,05,000 over six months, so the programme needs a 2.5 point lift, not two, to break even. Now you know what to watch in the month after the workshop, and you can drop the programme early instead of arguing about it in October.

Where this stops working

Some programmes have no countable unit. Leadership development, compliance training, a first-time manager programme: there is no order, no rework hour, no collection. Forcing a rupee value onto those produces exactly the fiction this article is about.

For that work, use a return on expectations instead. Kirkpatrick Partners' position is that in complex systems results are rarely caused by a single programme, and that stakeholders should define success in advance and be measured against it. That means a written expectation from the sponsor before the programme: what six people should be doing differently by a named date, observed by named people. It is not ROI. It is honest, and it fits decisions the business actually makes.

The other limit is data. A trend line needs history, a control group needs a comparable group, manager estimates need managers who watch the work. If none of those exist, the correct output is "we cannot attribute this", followed by a plan to be able to attribute the next one. That answer is more useful than a fabricated percentage, and it is far more likely to get you the baseline you asked for.

The three lines for your next review

Before the next programme is booked, write the measurement plan as three sentences and get the sponsor to initial it:

1. Fully loaded cost, including participant hours at loaded salary.
2. The one business metric and the isolation method, named now, not in October.
3. The break-even attribution share, computed from those two, with the decision rule attached.

If the plan cannot produce those three lines, the programme is not ready to be measured, and it is probably not ready to be bought either. Run the assessment and reporting side on the same discipline and the numbers stop being a negotiation. The one-page impact report format is the place those three lines go, and setting the baseline before the first session is what makes them possible. Everything else in an ROI conversation is commentary.