A technical product manager at PERQ, a software company, uploads recorded engineering conversations to an AI tool when the terminology gets away from her, and asks it to summarize what she needs to do next. She runs emails, support tickets, and client meetings through the same process, including the meetings where clients describe what they want the company to build.
That is one named person at one named company, and the workflow is becoming unremarkable. Goldman Sachs launched its AI assistant firmwide in 2025 after 10,000 employees had already adopted it for document summarization and data analysis. Citi deployed tools that summarize and compare documents and search internal policy.
The accuracy figures are bad enough on their own. A peer-reviewed study testing AI-generated security incident summaries found that fully autonomous summaries omitted critical details in 35 percent of cases and introduced factual inaccuracies in 42 percent. Those summaries go to stakeholders, auditors, and legal teams who will never see the raw incident data. Downstream, the summary is the event.
But the accuracy of any given summary is only half the problem. The other half is whether anyone in the organization can still tell when a summary is wrong — and that ability depends on expertise maintained through practice.
Accuracy is a property of the output. The organization's capacity to evaluate that output depends on skills — and skills require practice that the AI's presence displaces.
Take a legal team that starts using AI to summarize contracts instead of reading them. At the outset, somebody on the team could still read the original if a summary looked wrong. The time savings are real and the expertise is intact. Over the following months, the people who used to read those contracts are assigned elsewhere. New hires learn the summarized workflow as the workflow. Knowledge of what to look for in a particular clause structure, of what a counterparty's drafting habits reveal about its intentions, lives in the doing of the work; routine reading was the training through which junior lawyers acquired the pattern recognition to notice what mattered. As the model becomes the default first reader, the ability to catch its errors rests on expertise that the model's presence is quietly thinning.
There is no clean documented case yet of an organization discovering that this capacity had eroded. What we have are episodes where verification was assumed to be happening and was not. An Alabama law firm submitted federal filings containing five nonexistent AI-generated citations; the division head told a judge he had scanned rather than reviewed the documents, on the understanding that a colleague had already checked them. Deloitte Australia refunded part of its fee for a 237-page government report after an outside researcher found fabricated sources and a misquoted court judgment, errors that had survived every internal review until someone external compared citations against originals.
Those are verification failures rather than evidence of atrophy. But they point in the same direction. When the model reads first and the people who could read the original are doing other things, the distance between summary and source widens, and nothing in the organization is measuring that distance.
The accessibility community, as the companion piece describes, built its verification instincts out of years of depending on intermediaries. Enterprises are acquiring the dependency without the history. The incentive structure explains why: every hour an employee spends on the original instead of the summary is an hour the tool was purchased to save. The economics of deployment run directly against the practice of checking. And the cost of that trade stays invisible until someone finally goes back to the source document and finds that the people who knew what to look for in it have been reassigned.