Data Depreciation: Why Stale Records Destroy AI Value and How to Add Shelf‑Lives

A Harvard Business School study found that 100 MB of text data loses half its predictive value after seven years, and the depreciation curve is non‑linear, meaning stale data cannot be compensated by volume. In AI systems, stale data produces fluent but incorrect answers, a silent failure mode called freshness rot that can go unnoticed until a downstream error occurs. Claude Code’s Auto Dream feature demonstrates the need for memory maintenance, consolidating 913 sessions in under nine minutes to prevent contradictory or outdated knowledge from misleading the agent. Batch‑processing pipelines that wait for nightly rebuilds create staleness windows of 18 hours or more, destroying business value with each hour of latency. Gartner reports that without metadata, data quality, and observability, more than 60 % of AI projects fail to deliver on business goals. The article argues that production data pipelines must assign explicit expiration dates and freshness SLOs to each record, mirroring the time‑value‑of‑money concept used in finance. By processing data where it is generated and moving only the results, organizations can avoid the depreciation curve and maintain model accuracy. The lack of a shelf‑life discipline in data pipelines is a systemic risk that scales with every AI decision made.