Stop Chasing AI ROI; Start Defining Real Value

Companies complain they can't see returns on AI. I'll be honest — most of them aren't looking in the right places.

The Finance & Commerce piece is right to flag a gap between investment and visible payoff. But here's the thing: that gap is often created by accounting and expectation, not by the algorithms themselves. Boards demand neat financial boxes — capital expenditure here, operating expense there — and executives dutifully paste AI projects into those boxes. The result is a pile of pilots that look expensive and diffuse on a spreadsheet because the real value lives in processes, optionality, and speed, not in one-off revenue lifts you can attribute neatly to a model.

Short-term ROI metrics are seductive. They're clean. They flatter quarterly reporting cycles. They also miss where AI often pays off: fewer exceptions, faster decisions, better allocation of human attention. Those are economic benefits that drip slowly into margins and customer satisfaction; they don't always generate a dedicated revenue line you can circle in green marker.

Treating AI like a widget you buy and plug in will produce the kind of disappointment that reads like failure.

Measurement design matters. Companies need to distinguish between direct revenue impact (rare, but real) and indirect operational gains (common, but diffuse). That means new KPIs — cycle-time reductions, error-rate declines, employee time reclaimed — and experiments designed to capture causality over months, not just weeks. It also means governance: who owns the metric, who updates the baseline, who decides when a model is “good enough” to change how work actually happens. If you leave those questions to an overworked finance team and a vendor slide deck, you'll end up with the “no visible returns” headline Finance & Commerce reported.

Look back at the first wave of corporate IT. When companies rolled out early ERPs, lots of them grumbled that they couldn't see payback either. The ones that eventually did weren’t the ones who bought the fanciest system; they were the ones who used the rollout as an excuse to standardize processes, clean up data, and kill a few sacred-cow workflows. AI is replaying that story. The tech is only half the plot; organizational rewiring is the other half.

Which brings us to the hidden work.

Most organizations underestimate the implementation cost. Building models is the flashy part; boring work pays. Data plumbing. Model monitoring. Feedback loops to fix drift. Change management. Those things are neither glamorous nor cheap. They also determine whether a deployed model becomes a Netflix-recommendations-style win or a dusty dashboard nobody trusts.

Talent matters, but not in the way press releases suggest. Hiring a few data scientists won't cut it. Companies need product-minded engineers, domain-savvy analysts, and managers who can shepherd AI outputs into operational decision-making. That requires a rewire of org charts and budgets. It also requires admitting that many current IT and procurement processes are obstacles — legacy procurement cycles, rigid SLAs, and security rules that were not designed for iterative model development. Finance & Commerce hints at this friction; it deserves center stage.

There's another blind spot: data quality and governance. If your training data reflects upstream manual errors or policy quirks, the model learns the noise. That noise becomes business risk. Fixing it is a governance project. That governance project is where legal, compliance, and operations intersect — and where silos die. Boards hate messy cross-functional debates. But messy cross-functional debates are exactly the work that turns AI from a novelty into a recurring dividend.

Sure, but some companies genuinely do see quick returns. Low-hanging-fruit use cases exist — process automation, routing decisions, fraud scoring — and when a use case matches a clean dataset plus a high-volume, repetitive decision, returns can be immediate. The headline risk is that those wins get treated as the norm instead of the edge case.

Fast returns tend to show up where organizations had already solved the plumbing and designed the decision around automation. The “AI payoff” is really the payoff from years of process discipline meeting a new tool. If leaders interpret those examples as evidence that AI itself is an instant profit machine, they'll scale prematurely and replicate the costly failures the Finance & Commerce piece documents.

You can already see the split in the wild. Amazon and UPS didn’t start with “AI initiatives”; they started with maniacal focus on operations and data, then let machine learning quietly seep into routing, forecasting, and inventory. The returns look boring on paper — fewer miles driven, fewer stockouts — but they compound. Meanwhile, plenty of firms bolt a model onto a messy workflow and act surprised when nothing material changes except the cloud bill.

Look — Kevin Kelly likes to say that tools amplify existing behaviors; William Gibson wrote about a neon mirage of cyberspace that promised more than it delivered. Companies are living that tension now. AI will amplify whatever is already there: disciplined measurement and boring excellence, or scattered pilots and PowerPoint bravado.

Boards should start asking to see implementation roadmaps next to vendor demos. CFOs should treat data hygiene like infrastructure, not like a rounding-error experiment. CEOs should ask where, exactly, decisions will be different after deployment — and who owns making sure those decisions actually change.

If they don't, Finance & Commerce won’t need a fresh headline next year; they can just swap in a new example and hit republish.