The Captive Annotator: Usage Data, Service-Improvement Clauses, and the Uncompensated Expert Labor of Enterprise AI

Abstract: Companies buy workplace AI to cut professional drafting time, yet the software operates by recording how professionals correct its mistakes. Enterprise contracts guarantee that customer documents, prompts, and outputs will not train vendor foundation models. In practice, standard terms reclassify behavioral interaction as service data or usage telemetry, reserving rights to improve products and develop new commercial features. Through this contractual partition, the lawyers, clinicians, analysts, and engineers using the software become captive annotators, refining tools they do not own. While frontier developers pay outside specialists premium rates for this exact judgment, enterprise customers supply it as an unpriced byproduct of ordinary work. Contracts drafted by the vendor draw the line between maintaining a service and training a model. The customer finances the subscription and the salaries of the specialists who correct the drafts, only to purchase the improved system back. Keywords: Captive Annotators, Enterprise AI, Service Data, Usage Telemetry, Service-Improvement Clauses, Data Protection Addendum, RLHF, Direct Preference Optimization, Model Collapse, Uncompensated Labor, Digital Archaeology, Data Enclosure, Professional Verification, Human Oversight, Federal Trade Commission Section 5, Commercial Surveillance, Knowledge Work Automation, Synthetic Data Degradation, Customer Data Partition, AI Procurement Contracts