Ask for the ROI of an AI initiative and the reflex is a savings number: hours cut, headcount avoided. The reflex has two defects. It undersells — most of what AI actually delivers inside companies is not cost removal but capacity, speed and quality, none of which appear as a reduced line in the budget. And it mis-selects: measured only in savings, the projects that win approval are the ones that remove work, while the ones that create capability — faster analysis, better decisions, capacity the company did not have — lose to projects that merely delete. Both defects are fixable with a method.
Every initiative earns through one of a few mechanisms: it removes work (automation of a task), it accelerates work (the same task, shorter cycle), it raises quality (fewer errors, better output, stronger service), or it adds capacity (work the company could not do at all before). Naming the mechanism before the launch is what makes measurement possible at all, because the mechanism dictates what to watch: a removal project shows up in hours, an acceleration project in cycle time, a quality project in error and rework rates, a capacity project in the output of the previously-impossible work. Teams that skip this step end up trying to prove value in a currency their project does not emit.
ROI has a denominator nobody mentions in the enthusiasm: the honest «before». The hours a process consumed before automation, the cycle time before acceleration, the error rate before quality work — captured before the project starts, from actual records rather than optimistic recollection. Without a baseline the measurement becomes archaeology at the finish line, and archaeology always finds what the sponsor wanted found. The discipline is bookkeeping, not faith: the P&L effect of AI is legible only if someone wrote down what the world looked like before.
The uncomfortable half of the method is valuing quality, speed and capacity in terms an owner can weigh. Speed has a price when a cycle time has a business meaning: decisions made faster, clients served sooner, issues caught earlier — the mechanism states it, and the value follows from stated facts rather than invented multipliers. Capacity is valued by what the new capacity is spent on: the analysis finally done, the segment finally served. If the freed capacity has no destination, its value is honestly zero — which is itself a finding about the initiative, not the measurement. And quality is valued by what errors cost: rework, lost trust, warranty exposure — costs the company already knows from experience. None of these are savings in the accounting sense; all of them are real, and where AI creates value is precisely the territory where this valuation work pays.
A final peculiarity: AI initiatives change as they run — usage grows, the model's work shifts, the process around it adapts. A single post-hoc ROI calculation photographs an arbitrary moment. The working practice is checkpoint measurement: baseline at launch, first read when usage stabilizes, then on a cadence tied to the cost of the process — because the honest ROI of an AI project includes the project's own running cost, which is the cost most frequently forgotten after the approval.
Savings prove an AI project was cheap. Only the mechanism, the baseline and the destination of freed capacity prove it was worth doing.
Setting the measurement frame — mechanism, baseline, valuation of non-savings benefits, checkpoints — is standard work in the AI practice, done at launch because at the finish line it can no longer be done honestly.
We use cookies to analyse traffic and improve the site. Analytics is enabled only with your consent. Cookie Policy