The purge script runs nightly against the EDRMS without checking which records the AI layer still references through open queries. A faculty grant application from 2019 vanishes from the system even though the assistant pulls it daily to cross-check budget lines.
Developers built the integration assuming every document ID would remain stable for the life of the project. Instead the retention schedule treats every closed file as disposable once the regulatory window closes.
Now the assistant fabricates missing context from earlier snapshots stored in the vector index. Grant officers accept the output because it looks complete on the dashboard.
The EDRMS vendor sold the university a clean compliance story. They never mentioned that API consumers inherit the same deletion triggers without separate versioning endpoints.
Your records manager flags the mismatch in the monthly steering committee. Procurement still treats the platform as a solved records problem rather than an active data source for downstream systems.
Fixing it requires either freezing retention for AI-linked folders or building a shadow index that survives the purge. Both options sit outside the current support contract.
The pattern repeats whenever Australian institutions bolt AI onto legacy records systems without rewriting the retention logic first.