Ask a chief executive what their company did with AI agents last year and 97 percent will say they deployed them. Ask the people actually doing the job and the number drops to 52 percent. Ask the executives a second, more honest question, whether the strategy behind all that deployment is real or theater, and 75 percent say it is "more for show" than actual internal guidance. That is the finding of a survey from Writer, an enterprise AI vendor, run with Workplace Intelligence, a research firm: 2,400 respondents across six countries, fielded this past December and January, split evenly between C-suite and employees. The people who signed the purchase orders are the same people telling pollsters the whole thing is a performance, and underneath that performance sit the workers whose jobs are quietly being redrawn around tools nobody can yet prove are paying for themselves.
The performance has a budget. Fifty-nine percent of companies now spend more than $1 million a year on AI technology. Ninety-seven percent of executives call it beneficial. And yet only 29 percent report significant return from generative AI, and just 23 percent from AI agents specifically, the product category everyone spent last year deploying. Thirty-nine percent of companies have no formal plan to turn any of it into revenue. Forty-eight percent call the whole adoption effort a "massive disappointment." Writer has every incentive to sell the next round of tools, and that should color how much weight you give the more dramatic numbers in its own survey. But a vendor has no obvious incentive to publish that three-quarters of its own customers think their AI strategy is theater. When the salesman says the show is fake, believe the salesman.
The bigger, independent study makes the same point with less flattering framing. MIT's Project NANDA, in its July 2025 report "The GenAI Divide: State of AI in Business 2025," reviewed more than 300 publicly disclosed enterprise AI deployments and interviewed executives across industries. Lead author Aditya Challapally put the headline finding to Fortune bluntly: roughly 5 percent of enterprise generative AI pilots produce rapid, measurable revenue gains. The other 95 percent, representing an estimated $30 billion to $40 billion in spending, show no measurable impact on the bottom line. "The 95% failure rate for enterprise AI solutions represents the clearest manifestation of the GenAI Divide," Challapally said. Separately, S&P Global Market Intelligence's Voice of the Enterprise survey, covering just over a thousand IT and business professionals, found the share of companies abandoning most of their AI initiatives before they ever reached production jumped from 17 percent to 42 percent in a single year. Two different research shops, two different methodologies, the same shape: adoption everywhere, delivery almost nowhere.
Here is the part worth sitting with, because it is where the theater does the most damage. More than half of generative AI budgets are reported to go to sales and marketing tools, the functions a board can see and an earnings call can mention, according to independent summaries of the NANDA findings; the report's own document was not readable in full, so this figure and the two below rest on secondary accounts that agree with each other rather than a direct reading of the primary text. The actual return, where those accounts describe NANDA's researchers finding one, sat somewhere far less visible: back-office automation. Replacing outsourced customer support and document review, cutting agency spend, tightening procurement and finance workflows. Those accounts cite savings of $2 million to $10 million a year from automating exactly the kind of grinding operational work that never makes a keynote slide, alongside roughly a 30 percent cut in outside marketing agency costs. Purchased, vendor-built tools, deployed by companies willing to demand real customization the way you would from a business process outsourcer, a firm a company pays to run its back-office work, succeeded about 67 percent of the time. Companies that insisted on building their own succeeded roughly a third as often. "Almost everywhere we went, enterprises were trying to build their own tool," Challapally told Fortune, describing an instinct the data says is mostly vanity.
Coding is the case that should give the more triumphant version of this story some pause. It gets cited as one of the places AI agents unambiguously earn their keep. The most careful controlled test of that claim says otherwise, at least for the workers who know their jobs best. METR, a nonprofit AI research organization, ran a randomized trial in 2025 with 16 experienced open-source developers, an average of five years on their own mature codebases, doing 246 real tasks, half assigned AI tools and half not. The developers using AI tools were about 19 percent slower. Before the study, they predicted AI would cut their time by 24 percent. After finishing, having just been measurably slower, they still estimated AI had made them 20 percent faster. Economists and machine learning researchers surveyed beforehand predicted even larger gains. Everyone, including the people running the experiment, expected the tools to win. The tools that suggested code fastest were least useful to the developers who already held the codebase in their heads, because a fast wrong guess costs more to check than a slow right one. METR flags the result as historical now, tied to early-2025 tools on unusually demanding projects, and that caveat is fair. It is also the only piece of this discourse built on a randomized trial rather than a survey of people describing their own performance, which makes it worth more, not less.
Put the pieces together and the shape of the story is not "AI works" or "AI doesn't work." It is that AI mostly works exactly where nobody is watching, and mostly gets bought for exactly where everybody is watching. The back office absorbs the automation and quietly stops needing an outsourcing contract. The boardroom absorbs a sales and marketing tool it can announce on an earnings call, whether or not the tool does anything, because 75 percent of the executives who bought it already know it is mostly for show. Writer's own survey found 92 percent of C-suite respondents are cultivating a new tier of "AI elite" employees while 60 percent say they plan to lay off workers who do not adopt the tools fast enough. That is the part that should worry the person actually doing the job, not the shareholders. The theater has an audience, and it is not accountable to them. The bill, when a pilot quietly dies without anyone in the C-suite losing a bonus over it, tends to land on whoever's role got restructured around a tool that was never going to pay for itself. Ask who benefits from the announcement. Ask who carries the risk when the number turns out to be 5 percent. Ask who gets to walk away clean when the pilot folds. It is rarely the same person twice.
