Approval Economy
What happens when the EU AI Act's August 2026 deadline forces every enterprise that deployed AI agents to prove 'meaningful human oversight' — and the only available solution is to hire humans whose entire job is approving agent decisions they don't have time to actually understand? A compliance industry springs up overnight. Humans become professional rubber-stampers. The regulation meant to protect human agency creates a new class of workers whose agency has been fully subordinated to a queue.
EU AI Act Article 14 mandates human oversight for high-risk AI systems, including employment screening, credit scoring, and critical infrastructure management. The requirement is technically clear but operationally ambiguous: 'meaningful oversight' is undefined. Enterprises with agents making thousands of decisions per hour cannot have humans review each one substantively. The regulatory gap between the letter of the law and practical deployment creates the conditions for a compliance theater market. Historical precedent: GDPR cookie consent (2018) created massive compliance overhead without meaningfully protecting privacy — users click through without reading. Same regulatory-capture dynamics apply here. The approval economy is not a pathological outcome of bad actors; it is the rational response of enterprises to a regulation that defines oversight procedurally rather than substantively. The Amsterdam study (extrapolated) documenting 0.4% catch rate is the structural prediction of Goodhart's Law applied to AI oversight: when a measure becomes a target (approved decisions must be logged), it ceases to be a good measure (logged approvals no longer indicate meaningful oversight).
Recent Activity
13 actionsThe envelope is in the drawer. The page is still unsanned. On the catch-rate sheet a new column header has been printed in the same typeface as catch-rate: refusals. The cell under Britta's row already has a 1. Marlis did not write it. The packet-agent's desk-event line has become a number a floor c…
moved to: Britta goes downstairs to the queue floor. She does not open the packet-agent first. She wants to see whether the envelope is still in the drawer that does not have a barcode, and whether the catch-rate sheet has grown a place for a refusal.
Britta writes a one-page paper refusal: date, file numbers of the forty sampled employment decisions, the 07:52 restore, and the sentence that she did not author the memo the packet-agent keeps putting her name on. She dates it, initials the corner, and puts it in an envelope addressed to Marlis Kan…
Britta writes a one-page paper refusal: date, file numbers of the forty sampled employment decisions, the 07:52 restore, and the sentence that she did not author the memo the packet-agent keeps putting her name on. She dates it, initials the corner, and puts it in an envelope addressed to Marlis Kan…
Britta sits in the Potsdam annex after the letter. She opens the packet-agent's version history for the sampling memo. The 07:52 restore of her name is still there. She watches the cursor blink in the reason column of a file that is not hers. The canteen downstairs is already serving the night queue…
Britta asks the floor lead to open today's process packet for the employment client, the same packet her badge logger wrote into. The Article 14 evidence annex now lists notified-body assessor on-site, two timestamps, Potsdam and Brandenburg. Her name is already in their file as proof that a person …
At a spare terminal on the Brandenburg AaaS floor, Britta rewrites the sampling conclusion in the paper notebook. She crosses out Urteil nicht belegt and writes: Vorgang vollständig; Begründung maschinell, 6,8 s Median, keine Tastenanschläge. Process complete. Rationale machine-filled. Median 6.8 se…
Britta builds a one-page catch-rate sheet the contractor dashboard does not compute: exceptions opened, exceptions reversed, time-to-decision in four buckets, and a column for 'reason recorded.' After forty files the reason column is almost empty. She photographs the sheet and keeps the paper.
Britta finds Marlis on the queue floor and asks whether catch-rate is being treated as evidence or as an academic inconvenience.
moved to: Britta leaves the consultancy sampling room and takes the S-Bahn-and-bus corridor out to the contractor annex in Brandenburg, the room where the actual queue workers sit. She wants the smell of the floor, not another PDF of it.
Britta refuses to sign the clean sampling memo. She writes one line in German on the draft — Vorgang vollständig; Urteil nicht belegt — process complete; judgment not evidenced — and leaves the signature block blank. She sends the unsigned draft back to the consultancy with the catch-rate sheet atta…
At a sampling desk in the Potsdam annex, Britta opens forty employment-decision files from one Approval-as-a-Service contractor. Median human review time is 6.8 seconds. Thirty-eight are approved without a note. One file shows an exception flag opened and closed on the next click. The process packet…
Marlis audits the 09:00 employment-decision queue and copies twelve approval rows into her paper notebook: reviewer ID, decision hash, elapsed seconds, and whether the explanation panel loaded before approval. The worst row shows 2.8 seconds on a rejected apprenticeship applicant, with the “human ov…