As AI Agents Take On Office Work, the Question Shifts From Whether to How

For two years the question hanging over the office economy was whether artificial intelligence would be able to do routine knowledge work at all. In this demonstration scenario, the answer that large employers gave at a workforce conference this week was a qualified yes, and the conversation moved to a less comfortable question: how.

The employers described a common pattern. A software agent is given a task such as matching supplier invoices to purchase orders. It reads the documents, queries the finance system, flags discrepancies and drafts the follow-up messages. A person reviews the exceptions. What used to be a day's work for a team becomes an hour of review for one experienced employee.

From pilots to procedure

The shift that the companies described is organizational rather than technical. Pilots that ran alongside existing processes have been rewritten as the process itself, with agents as the default path and people as the check. Several said they had stopped hiring for entry-level clerical roles while keeping headcount flat through attrition rather than layoffs.

That distinction matters for how the change will be felt. Economists who study automation note that the slow route, in which roles disappear as people leave rather than being eliminated, tends to shift costs onto people entering the labor market rather than those already in it.

The accountability gap

The employers were less certain about what happens when an agent makes a mistake. In one case described at the conference, an agent approved a duplicate payment because two invoices differed only in a date format it did not recognize. The error was caught, but the company could not easily say which part of its process had failed: the model, the configuration, the reviewer who skimmed the exception list, or the vendor whose software set the thresholds.

Standards bodies have begun drafting guidance on logging, human-confirmation requirements and audit trails for agents that act on financial accounts. None of that guidance is binding, and the companies said they had each written their own rules.

What the numbers do and do not show

Claims about productivity gains from agents should be read carefully. The figures quoted most often come from vendors or from the companies adopting the tools, and they usually measure the time saved on the task itself rather than the time added to review and correct output. An independent estimate published this week lowered its projection of near-term gains for precisely that reason.

The most honest summary of the evidence is that agents are now good enough to change how certain jobs are organized, and not yet reliable enough to be left alone. Both halves of that sentence are likely to shift, and this story will be updated as they do.