The AI ROI Calculation Founders Keep Getting Wrong

A founder sits across from a vendor demo and hears that the tool will "improve decision quality" and "increase team alignment." Both claims are plausible. Neither appears anywhere in a profit and loss statement. So the founder passes, files it under "nice to have," and moves on. The tool gets rejected not because it failed to deliver value, but because the value it promised lived in a column that finance doesn't read.

The problem isn't skepticism, it's category error

Iternal draws a clean line between two types of AI benefits. Hard benefits flow directly into financial statements: reclaimed labour hours converted to headcount reduction, eliminated tool subscriptions, fewer escalations to expensive specialists, avoided overtime. Soft benefits improve capability without immediate P&L visibility: higher output quality, faster cycle times not yet attached to a headcount change, better compliance posture, what Iternal calls a "trust dividend" when outputs need less rework.

Both categories are real. The mistake is presenting them as the same thing.

When a founder bundles "improved morale" and "three hours saved per week per employee" into a single ROI number, the finance leader reading it cannot tell which part is cash and which part is aspiration. The whole number becomes suspect. Not because the soft benefits are fake, but because they contaminate the hard ones.

What the formula actually requires

Wave and Symestic both push toward the same minimum standard: hours saved multiplied by loaded labour cost, minus tool cost, equals net gain. That's the floor. It's not elegant. It doesn't capture everything. It does survive contact with a CFO.

The loaded labour cost matters more than most founders realise. Symestic notes that investment cost includes not just the vendor quote but implementation time, internal project hours costed at a loaded rate, training, integration, and first-year operating overhead. Wave adds that hours saved only convert to cash when they attach to headcount moves or revenue-bearing work. Reclaimed time that gets absorbed into longer lunch breaks doesn't appear in the P&L either.

So the formula has a prerequisite: you need to know where the reclaimed time goes. If a three-person team saves two hours each per week and those hours redirect to client-facing work at a billable rate, the math is clean. If those hours disappear into Slack, the formula returns zero even though the tool worked.

When the hours-saved number passes but the investment still fails

Bloomfire identifies a category of hard ROI benefits that the simple formula misses: gains from faster onboarding, reduced turnover, fewer support escalations. These are quantifiable and carry specific financial values. They don't appear in a weekly hours-saved count.

The IT productivity paradox literature makes a sharper version of this argument. Technology investments that looked marginal on short-term output metrics produced significant gains when measured over longer periods and across quality dimensions. Benefits showed up through quality and variety rather than raw output. A founder applying a hard-only formula to a knowledge management tool in month three will likely see a marginal number, not because the tool is failing, but because the formula excludes the dimension where the tool performs.

This is a real objection. A calculation that structurally excludes quality gains doesn't produce a conservative estimate. It produces a wrong one.

The problem is that including soft benefits in the headline number doesn't fix this. Mixing them in is precisely what causes finance leaders to reject otherwise sound investments. The research is explicit on this point. The fix isn't to remove soft benefits from the conversation. It's to stop presenting them as if they were already in the P&L when they aren't there yet. Report them separately, with the evidence standard each deserves, and let the hard number stand on its own.

What the minimum proof of value looks like

McKinsey's 2023 estimate puts generative AI's potential annual value at between $2.6 trillion and $4.4 trillion across 63 use cases. That number is real and it is also useless to a founder evaluating a $400-per-month writing tool. The aggregate figure incorporates long-term organisational redeployment across entire sectors. It doesn't translate to a quarterly tool decision.

What does translate: take the loaded hourly cost for the role using the tool, multiply by hours saved per week, multiply by weeks in the period, subtract total implementation and subscription cost including your own team's setup time. If that number is positive, you have a floor. Present the soft benefits below it, labeled clearly, with whatever survey or outcome data you have.

The founders who reject AI tools aren't wrong to demand proof. They're wrong about what counts as proof.