Trust in the Account: Visual Evidence Cues and Appropriate Reliance on AI-Prepared Administrative Decisions

When an artificial-intelligence system prepares an administrative action on a citizen, the official who must stand behind that action does not — and cannot — inspect the system. They inspect the account the system gives of what it did: the claims, sources, timings, and policy bases assembled into a case dossier. Yet the dominant interface cue for that account is a single green "verified ✓" badge, and the dominant human-factors result is that such confident reassurance is enough to capture the reviewer. Automation bias and complacency, the repeated failure of explanations to correct over-reliance, and the documented hollowness of mandated human oversight together imply that Article 14 of the EU AI Act — which requires a natural person to oversee high-risk AI — is an empty duty unless the official can actually read the account. This paper makes a conceptual move and develops its consequences for interface design. The move is to relocate trust's object: from the system (a person's standing attitude toward a technology, as measured by the Human-Computer Trust lineage) to the account (a specific, inspectable artifact of action). We call this Trust in the Account. From it we derive (1) a catalogue of visual evidence cues — provenance, freshness, scope, conflict, policy basis, contestability, and independent verifiability — each answering one reviewer question that the "verified ✓" badge silently swallows; (2) an Attestable Account Workspace, a four-column reviewer interface that closes Norman's gulfs of evaluation and execution and supports situation awareness; (3) a Trust-in-Account instrument that re-stems the four Human-Computer Trust dimensions onto reliance on an account, with a pre-registered prediction that a general trust score stays flat across record states while Trust-in-Account tracks record quality; and (4) a taxonomy of record states — sound, stale, replay/out-of-scope, misleading-"verified" — against which appropriate reliance is defined as discrimination, not blanket distrust. We close with a research agenda for a synthetic-lab programme using synthetic stimuli and online proxy participants. The contribution is a framework for attestable trust and a set of transferable, buildable interface-design principles for operationalising human oversight of AI-prepared public-service decisions.