The Number That Wins Isn’t Always the Number You Fund

Give a finance team two capital projects and a spreadsheet, and they will hand you back a winner within the hour. NPV, IRR, payback: run the formulas, rank the outputs, fund the top of the list. It feels like a decision. Most of the time it is closer to a coin toss dressed up in decimal points.

But this doesn’t usually show on the board deck: on a real capital decision, NPV and IRR frequently disagree with each other, and the discount rate you picked to run the numbers can flip the ranking entirely. Deloitte’s Finance Trends 2026 report, surveying more than 1,300 global finance leaders, found that 30% cite the need to bolster advanced scenario-planning capabilities specifically to manage this kind of uncertainty. Roughly a third of finance leaders, on the record, admitting that a single-point estimate is not enough. Finance teams know how to compute a metric. What they don’t have is one number that actually settles the argument, and most of them have known that for years.

I want to walk through why that happens, using real numbers, and then show you the workflow I built to actually work the problem instead of admiring a ranked list.

The two-project problem, honestly stated

Picture a mid-size manufacturer sitting on a big signal: a large new customer wants a lot more volume than the current plant can produce. The board has one capital budget and two mutually exclusive ways to serve it. Expand the existing site, fast and familiar. Or build a second site, slower, pricier, but built exactly for the new demand and cheaper to run once it’s live.

Every capital committee has lived a version of this. And the textbook answer, “just compute NPV and take the higher one,” breaks down the moment you actually run both projects side by side, because the two metrics that matter most almost never agree.

NPV tells you the total value created, in today’s dollars, at your cost of capital. IRR tells you the rate of return the project itself generates. They are different questions. A project can create more absolute value while returning a lower rate, if it is bigger and takes longer to pay back. That is not a modeling error. That is what capital intensity does to a return profile.

The worked case

I built this on a synthetic beverage manufacturer, Silverbrook Beverages: a $48M-revenue producer running near capacity, with a letter of intent from a major retail chain for roughly 45 to 70 million cans a year starting in two years. The board has four weeks and one capital budget of $15M. Two mutually exclusive options are on the table.

Same evaluation policy for both: incremental after-tax free cash flows only, an eight-year horizon, a 9% discount rate, a 12% hurdle rate for execution risk, 25% tax, working capital built at 10% of incremental revenue and recovered at the end. No blending the two into one fake combined score. Financial results and strategic scores sit side by side; weighing them is the board’s job.

Run both projects through that policy and here is what comes back:

That table is the trap. Option B wins on NPV. Option A wins on everything else: IRR, payback, and profitability index. Neither project dominates the other. If your process stops at “compute NPV, fund the winner,” you fund Option B and never learn that Option A was the faster, higher-return, better-per-dollar-invested way to capture nearly the same value.

Where the ranking flips

Push the discount rate up and the story changes again. At 9%, Option B leads by roughly $1.0M of NPV. Run the same cash flows at 12%, and the two projects are within about $17,000 of each other, essentially tied. Above roughly 12%, Option A pulls ahead and stays ahead as the rate keeps climbing.

That crossover point is not a rounding curiosity. It sits almost exactly at the board’s own 12% hurdle rate for execution-risk projects, which means the ranking genuinely depends on how much execution risk you believe Option B, the newer site with a new team and new permits, actually carries. Change your belief about risk by a point or two, and you change the winner. A single-point NPV printout hides that entirely. Nobody who looked only at the 9% number would ever know how close, and how conditional, this decision really is.

The lesson generalizes past this one case: a base-case NPV is a single draw from a distribution of possible outcomes, not the outcome.

Then the downside case, and the near-tie that actually matters

The letter of intent behind this decision is real interest, not a signed contract. So the honest next question is: what happens if the customer delivers less volume than promised, later than promised, at tighter pricing than promised? I ran three scenarios, weighted by probability: a 30% chance of a partial, softer delivery of the deal; a 50% base case where it lands as signed; a 20% upside where the relationship expands further.

In the downside case, Option B’s NPV does not just shrink. It goes negative, to roughly minus $2.8M, because the second site was purpose-built for volume that did not fully show up, while Option A, sized more conservatively, stays comfortably positive at around $1.4M. In the upside case the pattern reverses hard, and B pulls ahead by more than $4.5M.

Blend all three scenarios by their probabilities and the probability-weighted expected NPV comes out to $3.97M for Option A and $4.11M for Option B, a gap of about $141,000 on projects worth $4M-plus each. That is, for practical purposes, a coin flip once risk is actually priced in.

That is a judgment call, and it was always going to be one, no matter how many decimal places the model returned.

For what it is worth, the tool’s own draft recommendation lands on Option A: fund the expansion now, and revisit the greenfield site the day the letter of intent converts to a binding contract. Notice the shape of that sentence. It is not “Option A is better.” It is a conditional call with a named trigger for reversing it. That is what a draft recommendation from a model should sound like: a starting position, stated with its own limits attached, waiting for a person to either sign it or argue with it using something the model could not see.

What I actually built: the workflow

I did not want a report that told me which number to trust. I wanted a workflow that showed me where the two projects genuinely disagreed and put the disagreement in front of a human before anything got built into a persuasive tool.

1. One skill, run twice

   The same investment-appraisal logic runs once per option, in parallel, never comparing them to each other. Each run states its formulas explicitly: free cash flow as after-tax EBIT plus depreciation, minus capex, minus the change in working capital, plus salvage. Each option gets its own NPV, IRR, payback, discounted payback, and profitability index, computed the same way, so the comparison that follows is apples to apples.

2. The stress test

   This is where the base case stops being treated as gospel. The workflow reruns both projects’ cash flows across the discount-rate curve to find the crossover rate, and across the three probability-weighted scenarios to find the expected NPV and the downside case. Nothing here decides anything. It just makes the fragility visible.

3. Two scorecards, never merged into one

   A financial scorecard sits next to a weighted strategic scorecard, six criteria the executive team scored by hand: capacity headroom, execution risk, speed to revenue, unit-cost competitiveness, flexibility, and fit with the new customer relationship. The strategic scores are set by people, not the model. The math only totals the weighted score. Option B wins the strategic scorecard clearly. Option A wins on speed and execution risk. The two are never blended into a single number, because blending them is exactly the move that would hide the trade-off instead of showing it.

4. The human in the loop, before the pretty tool gets built

   This is the step that matters most and the one most workflows skip. Before anything gets turned into a polished decision tool, the assumptions get challenged. Is the 12% hurdle rate actually right for a purpose-built greenfield site, or should execution risk be priced even higher? Does the downside scenario understate how a soft launch really behaves? What does the balance sheet’s financing headroom actually support before the board needs a term loan instead of the existing revolver? That conversation happens on the raw numbers, before a compelling chart can talk anyone out of asking it.

5. The decision tool, once the assumptions survive the challenge

   Only after that gets built: a single self-contained page with a discount-rate slider, scenario toggles, and a letter-of-intent delivery slider, so the board can move the assumptions themselves in the room and watch the ranking respond live, instead of trusting one static number on a slide.

Why the split is the whole point

The rules own the math. NPV, IRR, payback, the scenario-weighted expected value: those are formulas, computed the same way every time, and you do not want a probabilistic model doing that arithmetic by feel. The calculation layer stays deterministic.

The AI’s job is the analysis that used to be too expensive to do properly: running both options through every scenario, finding the exact crossover rate, holding two scorecards apart instead of collapsing them into a false consensus number, drafting the language that explains what moved and why.

The human’s job is the part no model gets to do: deciding how much to trust the letter of intent, how much execution risk a new site really carries, and which project the board can actually live with if the downside shows up. The AI made that decision possible to make well, in an afternoon, on real numbers instead of gut instinct. It never made the decision.

The number that wins on a single slide is not always the number you should fund, and the only way to know the difference is to make the model show you where it disagrees with itself before you ever open the deck.