Hard vs soft AI ROI: what actually counts

Almost every team has a soft AI ROI number and almost none has a hard one, and the gap between them is where AI budgets get killed. The clearest illustration came from a Gartner poll of CFOs: about 74% reported time savings from generative AI, but only 5% reported cost reductions and 6% saw revenue or profit lift. Time saved is the soft number nearly everyone has. A dollar on the P&L is the hard number almost nobody has. Hard AI ROI is the kind a controller can point to, labor cost removed, cycle time cut, error rate down, throughput up. Soft AI ROI is everything that's real but doesn't land in a financial line: hours back, better morale, higher-quality work, freed capacity. Both are genuine. Only one gets funded, and confusing them is how good AI work loses its budget.

Finance isn't being stubborn when it discounts soft ROI; it's being burned by experience. Soft numbers don't auto-convert into money, and a lot of them leak before they ever could. A Workday study found that roughly 37% of the time employees save with AI gets spent fixing the AI's output, so the headline hours overstate the recoverable ones. And even clean savings rarely become cost cuts on their own: an EY survey found only 17% of organizations with AI productivity gains reduced headcount, while most reinvested the time into more AI work or new output. That's often the right call, but it means the saving stays invisible on a cost line. The pattern is the same one that makes AI ROI hard to see in general: value gets felt, then dissipates before it reaches the P&L. Finance approves hard numbers because they survive that leak. It waves off soft ones because, on their own, they usually don't.

There's a serious argument that finance is the problem here, not the soft number. The strongest version says a quarterly financial lens fundamentally can't see how AI creates value in knowledge work: efficiency that compounds, fewer errors, retention, one person now doing specialist work, better strategic position. We get what we measure, the argument goes, so a company that demands quarterly profit will kill genuinely valuable AI work because its yardstick can't register it. That critique is right that soft value is real and that a quarterly-only lens destroys good work. Where it breaks is the conclusion. The answer isn't to stop discounting soft ROI and ask finance to take it on faith. It's that every dimension on that list can be converted into a defensible hard number, and handing over the soft number without doing the conversion is what leaves good work unfunded.

Hardening soft ROI is a specific, learnable move, and it's the payoff of the whole exercise. Tie saved hours to a decision: either they cut a real cost, get redeployed to revenue-generating work, or raise a measured output like invoices processed or tickets resolved per person. Hours that aren't tied to one of those three are slack, not value, so name which one applies. Convert quality into error rate times the cost per error. Convert retention into recruiting cost avoided per role, since turnover runs a real fraction of salary. None of this requires pretending soft value is hard; it requires doing the arithmetic that turns it into something a controller can book. That conversion only happens when adoption is deep enough to change a whole workflow rather than speed up a single task, and it lands cleanest in the roles whose hours map directly to output. Bring your CFO one hardened number per claim, defend the baseline and the attribution, and the soft-versus-hard argument disappears, because you've stopped having it.