Do AI systems reliably retrieve from disorganized material dumps?

Nielsen Norman Group's Tanner Kohler describes a coping pattern among participants who build agentic AI systems but won't or can't clean up the mess those systems accumulate: "when users can't or won't clean up a mess, they contain it: everything goes into one place so the rest of the system stays tidy." The excerpt is explicit that this containment is opaque to the user by design, not by accident: "in agentic systems, the rug is really a junk drawer — the user keeps needing what's in it, but only the AI can find anything." One participant illustrates the failure mode concretely: he "emailed himself articles and links throughout the day for the AI to collect," which piled up in "a folder called to be synthesized, with no naming scheme." When he later asked the AI to pull relevant material from that folder for a project, "it retrieved the wrong articles, and he spent 45 minutes fixing the instructions instead of doing real work."

Kohler frames the mechanism through a Harry Potter analogy: the user is like Harry visiting his Gringotts vault, who "knows he has a vault, and he vaguely knows what's in it, but he can't find it or access it without a goblin guide." The AI plays the goblin, navigating the pile "surprisingly well," which is precisely what "makes it tempting to keep sweeping more and more under the rug" rather than organizing it at the point of capture. Kohler treats the strategy's durability as unresolved rather than settled: "whether the AI can keep compensating as the pile grows is an open question; the 45 minutes lost to the wrong files suggest it has limits."

This is the specific coping mechanism that Does AI abundance undermine the engineering discipline that scarcity enforced? names only in general terms as bolting-on and sweeping under the rug — the junk drawer is what that bolt-on accretion looks like applied to unsorted personal material rather than to code or system structure. It also bears directly on Does human-in-the-loop design actually enable effective oversight?: Data & Society's split of oversight into knowledge, observation, control, and intervention describes exactly the gap the junk drawer produces, since the user retains no knowledge of the folder's contents and no observation of how the AI retrieves from it, delegating both to the AI by construction. And it is a concrete instance of Does granting agents more autonomy undermine human oversight?: the junk drawer is a case where a user voluntarily trades the positioning needed to oversee their own material for the AI's ability to navigate disorder on their behalf.

The excerpt reports one concrete failure, the 45 minutes lost to a wrong retrieval, but does not quantify how often containment strategies fail versus succeed, and does not test whether minimal naming or tagging at capture time would have prevented the error. Kohler poses the durability question as open rather than answering it. What the excerpt supports, at the strength it allows, is that containment defers the cost of organizing material rather than eliminating it, and that deferred cost resurfaces as retrieval failures whose size scales with how much has been swept under the rug.