Auditing AI agents across a model change
On November 17, Microsoft announced the AI Transparency Initiative with more than 40 organisations. It aims to develop an open approach to transparency, security and governance for AI models and agents. When a company switches models, the practical test is whether it can still explain a past decision and check what the new agent does.
1. What Microsoft announced
Microsoft plans to expand its participation in NIST’s agent standards initiative and incorporate the resulting work into Microsoft AI, Foundry and Copilot. The announcement outlines standards work and planned integrations. During procurement, request an actual export for the product, version and environment under consideration, then check that your audit tool can read it.
2. What this changes for Belgian and French companies
An SME or mid-market company that switches models to manage costs or bring processing closer to its data needs to be able to review decisions made under the previous version. Ask the process owner to reconstruct a test case from an export without access to the provider’s console. IT departments and public administrations can make this an exit requirement and check who can access the records and how long they are kept.
3. Underside analysis: trace a case from start to finish
For a RAG system preparing a purchase request in Odoo Enterprise, use a case identifier to connect the model version, references to retrieved documents, the identity used, the reviewer’s decision and the ERP draft ID. Retain the references and decisions needed for audit. Collect full text only where justified and restrict access to the logs. For a local model running on a managed Mac or a cloud service, check where the records are stored and who can export them.
4. Check that the audit trail survives the switch
Run three fictional cases on both versions: an accepted request, a document the agent cannot access and a draft the reviewer rejects. Check that each export identifies the model version, distinguishes a refusal from a technical error and lets you find the outcome in Odoo. If you can no longer reconstruct a past decision, defer the switch until the old evidence remains readable and the new export can be used.
Before switching, ask someone outside the project to review exports from both versions. They should be able to link each proposal, refusal and approval to its case.
Assess an AI agent’s audit trail