Keep the source data usable

A lot of AI workflows depend on material that was written for some other purpose. Documents are structured inconsistently, old versions are still searchable, permissions get inherited wrongly, and updates turn up with no metadata. Keeping all that in a usable state is part of running the workflow, so it goes in the operating model.

Set up a source register with an owner for each collection. For each one, record its authority, scope, how often it changes, access rules, retention, supported formats and whatever signal tells you a piece of material is current. For retrieval systems, record chunking and indexing status as derived state. A document being indexed doesn't tell you it's valid.

Source work includes things like:

• adding and classifying new material
• withdrawing or superseding old versions
• fixing parse and extraction failures
• reconciling documents that conflict
• testing permissions after organisational changes
• rebuilding indexes after a material change to processing
• checking that citations resolve to the version that was actually used
• answering owner questions when the source itself is ambiguous

Decide what happens when you can't tell whether a source is fresh. The workflow might block the task, warn the reviewer or limit what it outputs. If it carries on with an unknown version, that should be a risk decision somebody made on purpose, and you don't want it happening as a silent fallback.

I'd measure source failures separately from model failures. Record whether the source was current, authoritative and available to that user. Otherwise a perfectly good answer grounded in an obsolete document can get blamed on the prompt or the model.

Budget source upkeep per collection, based on how often it changes. A stable set of approved forms needs much less than procedures that several teams keep updating. And include the people who understand the material, as well as the engineers running the index.