A study of consumer lending in Germany records a case anyone who has filled out an application will recognize. Parents applying for a housing loan were on parental leave. The online form had no way to represent that. The predefined fields offered, as the researchers put it, "no way to enter a free text field." A branch employee saw the problem and moved the application into a manual process.
That employee was doing a specific kind of work: converting a circumstance the institution had not anticipated into something the institution could act on. The work does not disappear when an AI agent handles the intake. It happens faster, at volume, and without the discretion to route an odd case somewhere else.
Call it categorical mismatch: the gap between a person's situation and the fields available to describe it. It is old and pervasive. State unemployment systems verify income against quarterly employer wage records, which leaves freelancers and gig workers outside the data source entirely. Tenant screening companies market conventional credit scores even though, as the CFPB found, only about 2% of adult renters have rental-payment history in nationwide credit files. The landlord gets a detailed financial profile in which the single most relevant behavior is usually absent. The categories fit whichever population was most common, or easiest to measure, when the system was built. Everyone else gets approximated.
What makes agent mediation worth separating out is that the intermediary itself can change the profile an institution receives. A study of life insurance applications in North America compared applicants screened through financial advisers with applicants screened directly by the insurer's telephone operators. Those screened through advisers disclosed systematically fewer medical and lifestyle risks, and received more favorable terms. The screening channel, not the applicant's health, shaped what the insurer saw.
The researchers put this down to incentives and scrutiny. Insurer interviews were recorded and conducted by salaried operators with no stake in whether the policy sold; advisers had commissions and continuing client relationships tied to closing it. What survived the translation depended on who was doing it and what they stood to gain from the answer.
When the intermediary is software, the translation goes invisible in both directions. The applicant may never learn which of their circumstances mapped cleanly onto a field and which were dropped or rounded off. The institution has no way to know what it did not receive. The W3C/GS1 workshop this month took up data sharing, privacy, and consent in agent-mediated commerce; whether the version of a person that arrives through an agent resembles the version that would have arrived directly is a different question, and largely unexamined.
The circumstances least likely to have a matching field are the ones that were uncommon or inconvenient to measure when the intake system was designed: parental leave, irregular income, households that don't resolve into the available options. The branch employee who pushed the German application into manual review was exercising a modest and unglamorous form of institutional judgment. This doesn't fit the form. Let me find another way.
An agent that filters a person's reality into institutional categories before either party sees the result does not create the mismatch. It may remove the one participant with the discretion to notice it.