Enterprise AI has won the argument about whether to deploy and lost the one about whether it works. A new survey of the world's largest companies, Plug and Play's 2026 Enterprise AI Strategy Pulse, finds that 74 percent now run at least one AI system in production and 93 percent are piloting or further along. And yet half of those production-stage companies cannot say whether any of it has paid off. The deployment threshold has been crossed, in the words of one of the survey's partners, but the value threshold has not.
The numbers get stranger the closer you look. The narrowest, simplest deployments, a single AI use inside one business function, are the worst measured of all: three-quarters of those companies say return on investment is too early to track or is not tracked at all. That is backwards, since a single function should make cause and effect easiest to see. The likely culprit is timing. A use case that goes live without a baseline metric is unmeasurable forever, because there is nothing left to compare it against. Larger studies echo the gap. KPMG found only 7 percent of leaders report established AI returns; Deloitte found two-thirds see productivity gains but only a fifth see new revenue.
A recurring diagnosis across this week's writing is that AI does not fix a company's weaknesses, it exposes them. Jayasri Ranganathan, a technology strategy lead writing in Forbes, puts it plainly: AI is an amplifier, and in organisations with weak data, fuzzy processes and unclear ownership, what it amplifies is the visibility of every gap. Gartner reckons 63 percent of organisations lack, or are unsure they have, the data practices AI needs. The bottleneck, she argues, is almost never the model. It is inconsistent processes and poorly governed data that the model simply refuses to paper over.
The problem sharpens as companies hand real authority to AI agents that can move money, approve requests and change records without a human in the loop. In traditional software, governance was the last box to tick before launch. That sequencing collapses when the software makes probabilistic judgments, argues a piece in Observer: an agent needs the same things a new employee gets on day one, a verified identity, a defined scope of authority, and a log of everything it does. The companies scaling agents with confidence built those controls in from the start; the ones that bolted them on afterward are finding it slow and expensive.
Veeam's Rick Vanover frames the coming failure vividly. The next big AI incident, he writes, will not look like a dramatic cyberattack with a ransom note. It will look ordinary until it isn't: an agent quietly updating thousands of records at machine speed, with legitimate work tangled up in the errors, so that rolling back the damage means erasing good business along with the bad. His proposed fix, "precision rollback," reversing only the corrupted changes, depends on governance that was designed in advance. The through-line across every one of these reports is the same. The hard part of enterprise AI was never getting it running. It is proving, governing and, when necessary, undoing what it does.