Adoption Is Not Returns
Stop measuring your AI program by how much of it you have deployed. The money has turned, and this week the bill came due.
Axios reports that banks are pulling back as the costs of infrastructure and chips far outpace current revenue from AI services, with Wall Street increasingly vocal that AI spending may not generate returns for most companies. That is the whole story for the C-suite this week. The patient capital that funded two years of experiments is now asking what it bought. Leaders who can answer in dollars will keep their funding. Leaders who answer in adoption metrics will watch capital get scarcer and more expensive.
Adoption was never the proof
The habit under threat is counting usage and calling it progress. McKinsey makes the case plainly: many organizations track AI adoption metrics rather than business value, and adoption rates don’t prove ROI. Fortune sharpens the warning with history. Fifty years of digital innovation drove rapid adoption, yet during that same period productivity growth declined. The question now is whether AI will follow the same pattern.
It can. Speed of uptake has never guaranteed output. If your board update leads with seats, logins, and pilots launched, you are reporting the exact number that fifty years of evidence says is disconnected from growth. Swap it for the number financiers want, which is value captured per process.
Pilots lie about production
The second trap is believing a cheap pilot predicts a cheap system. It does not. McKinsey finds that moving an AI tool from pilot phase to full production can multiply its cost by nearly four times, and that pilot success doesn’t predict production economics. A demo that thrilled the steering committee can quadruple in cost the moment it has to run for real.
This is why the pullback stings. Firms budgeted for the pilot and are now meeting the production invoice at the worst possible moment. The fix is to price AI the way you price anything serious. McKinsey’s value curve exists to identify which work processes AI can automate most cost-effectively. Use it to kill the low-impact pilots before they ask for four times their budget.
Value has a physical address
Returns are not abstract. They live in named processes and real constraints. McKinsey reports that autonomous AI agents are changing the economics of software engineering, and Cognition’s Russell Kaplan argues the next frontier is extracting measurable business value from these agents, because faster code writing alone doesn’t guarantee ROI. Even the clearest win, cheaper code, only counts when it reaches the ledger.
Value also runs into physics. Axios notes that the surge in electricity demand from AI is driving investment into geothermal, which is winning support across the political spectrum. Your returns depend on power you may not have priced. And they depend on trust. Atlassian’s Mike Cannon-Brookes argues that deploying agents is an accountability problem, that agents need clear naming and a human accountable for their mistakes. An agent no one owns is a liability, not an asset.
The firms that get through this are not the ones with the most AI. They are the ones who can name three processes where AI moved a real number, show what production actually cost, and point to the human who owns each decision. Everyone else was counting logins. The money has stopped accepting that answer, and it will not start again.