Enterprise AI adoption reached an inflection point in 2025, with global spending on generative AI tools crossing $48 billion across Fortune 500 companies alone. By every measurable standard, this should have been the year that artificial intelligence began delivering on a decade of promises. We benchmarked 200 enterprise AI deployments across 14 industries to find out what actually happened. The short version: someone spent $2.4 million so a chatbot could tell employees where the bathroom is.
What the Dashboards Show vs. What the Floors Look Like
Of the 200 deployments we examined, 161 were described internally as "transformational" in at least one executive presentation. The word appeared in board decks, investor calls, quarterly reviews. It's a useful word because it doesn't require you to specify what, exactly, was transformed.
On the ground, the picture is more specific. Forty-three percent of deployments were some variation of a customer service chatbot. Of those, 74% still routed to a human agent within two exchanges. The AI handled greetings and asked clarifying questions. The human handled everything after. In internal metrics, the chatbot was credited with "resolving" interactions that it had, in reality, triaged. One retail company reported a 60% AI resolution rate. When we audited the logs, the chatbot's most common resolution was generating a ticket and emailing it to the same support team that existed before the deployment.
Fourteen percent of deployments involved document summarization. These performed measurably well in demos and measurably worse when confronted with the actual documents a company produces -- contracts with nested amendments, compliance filings with jurisdiction-specific footnotes, engineering specs written by three teams across two time zones. The summaries were serviceable for documents you didn't really need summarized and unreliable for documents you did.
The Invisible Correction Layer
The most consistent finding across all 200 deployments was the emergence of what we're calling the invisible correction layer: employees whose unacknowledged job is to fix what the AI produces before anyone important sees it.
At a mid-size insurance company in Ohio, a claims processing team of nine was reduced to six after an AI tool was introduced to handle initial assessments. Within four months, the six remaining employees were working eleven-hour days. The AI generated assessments quickly. It also generated assessments that transposed policy numbers, misclassified water damage as fire damage, and once approved a claim on a property that had been demolished in 2019. The team spent roughly 40% of their time correcting outputs and the other 60% doing the work the AI wasn't assigned to. Nobody told leadership. The team lead said, candidly, "They announced this at an all-hands. There's no version of this where I go back and say it doesn't work."
This pattern repeated everywhere. Marketing teams quietly rewriting AI-generated copy that was technically fluent and tonally wrong. Legal teams re-reviewing AI-flagged contract clauses because the model couldn't distinguish between standard indemnification language and an actual liability risk. Engineers debugging code suggestions that compiled cleanly and failed silently in production. The work didn't disappear. It just became invisible, because the people doing it had every incentive to stay quiet.
The Productivity Paradox, Again
Sixty-eight percent of the companies in our study reported productivity gains from their AI deployments. When we asked how those gains were measured, the most common answer was time-to-first-output -- how quickly the AI generated an initial draft, summary, or response. This metric is real. AI does produce first drafts faster than humans. It also measures the wrong thing.
Time-to-first-output ignores review cycles, correction passes, and the meetings held to discuss why the output didn't match expectations. At one financial services firm, an AI tool reduced report generation time from four hours to twenty minutes. The subsequent review and revision process, which hadn't existed before because a human wrote the report correctly the first time, added three hours. Net savings: forty minutes. Cost of the tool: $380,000 per year. Cost of the analyst it was supposed to replace: $95,000.
We found this arithmetic repeated with remarkable consistency. The savings are real but narrow. The costs are real but distributed across people's calendars in fifteen-minute increments that never show up in a line item. Nobody is tracking the hour an account manager spends each week rewording AI-generated client emails so they don't sound like they were written by a courteous stranger.
What Leadership Believes vs. What Slack Channels Say
We surveyed 340 executives and 1,200 individual contributors across the 200 companies. Eighty-two percent of executives rated their AI deployments as "meeting or exceeding expectations." Thirty-one percent of individual contributors agreed. This is not a communication gap. It's an incentive gap. The people who approved the budget need it to have worked. The people using the tool daily have no mechanism to say otherwise that doesn't implicate the decision-makers above them.
In anonymous responses, employees described their AI tools as "a confident intern who never learns from mistakes," "autocomplete with a marketing budget," and, memorably, "a very expensive way to produce a first draft I immediately delete." These aren't people who hate technology. Many of them use AI tools productively in their personal workflows. What they resent is the institutional pretense -- the requirement to treat a marginal tool as a revolution because someone in the C-suite said it was one.
Where This Leaves Us
There are genuine uses for AI in enterprise settings. Code completion tools measurably help experienced developers. Translation services have improved. Data tagging and classification work well when the categories are clean. These are real, useful, unglamorous applications that don't make for good keynote slides.
The gap isn't between AI and no-AI. It's between what was purchased and what was needed. Between the press release and the pull request. Between the executive summary and the eleven-hour day. Somewhere in every organization that deployed an AI tool last year, there is a person quietly doing the job the AI was supposed to do, and doing it well enough that nobody notices the AI isn't doing it at all.