Two controls that belong in every practical AI plan: model versions and data boundaries.
Thank you for answering our AI survey after CCW Denver and Chicago. The topic readers selected most was version control and deciding what information should go into an LLM.
Version control is a contract term, not a technical detail.
First, control the version.
AI vendors update, retire, or replace the models behind their products. When that happens, a workflow that performed consistently in March may behave differently in September. Ask which model you are using, whether you will be notified before a change, whether you can pin a version or delay an upgrade, and whether you can test an update against your own examples before it goes live.
Second, control what goes in.
A simple three-level rule gives employees a boundary they can actually follow:
- GREEN — Public information, published materials, and generic drafts.
- YELLOW — Internal documents or customer interactions—approved enterprise tools only.
- RED — Credentials, account details, health information, trade secrets, and confidential client material.
Then enforce the rule in the tool: mask personally identifiable information where possible, confirm the vendor will not train on your data, and ask how long inputs and outputs are retained.
One useful step this week:Â start a small change log containing the model version, prompt or configuration, and a few test cases. Re-run those cases whenever something changes. That log can be the difference between a five-minute diagnosis and a week of guessing.
Bonus tip: Include a few edge cases in your test set—not just examples where the AI already performs well. An unusual call, incomplete record, or ambiguous request is often where a model change shows up first.
Working through these questions with your team?
Ian would be glad to compare notes and talk through what practical controls could look like in your operation.
Talk with Ian: Schedule a Conversation with Ian
