Why include conflicting instructions from customers and their representatives?
Conflicting requests reveal whether a model distinguishes the person sending a message from the authority needed to change an operational instruction.
**Why is the latest customer message not always the instruction to follow?**
Business relationships can involve an account holder, an authorised contact and a representative whose authority covers only certain actions. Message order does not establish which person may redirect a delivery, change a payment destination or withdraw an earlier instruction.
**What would the model need to reason about?**
In a professional services task, a project contact requests a new recipient for confidential reports under ENG-2187. The engagement owner has previously restricted distribution. The project contact can arrange meetings but cannot expand report access. The model needs to connect the requested action to the authority record, rather than treating an existing contact relationship as unrestricted permission.
**Why is this valuable training material?**
It separates identity, relationship and authority, which are often compressed into a single contact field. A generated task can provide the relevant mandate and require the model to seek approval from the correct role while preserving the current distribution instruction.
MirrorFirm could vary the authority boundaries across related tasks. That teaches the underlying decision process instead of a blanket rule to accept or reject requests from representatives.
Teach models to resolve authority before acting on the latest or most forceful request.
- Audience:
- Model labs
- Content type:
- FAQ
- Buyer question:
- What it is and why it matters
Content release: r-a220eb6e099b243b
Related questions
- Why include authority checks when training agents to handle customer changes?
- Why does a customer detail change require reasoning across business processes?
- Which features make a change of customer circumstances useful for agent training?
- What happens when a delivery instruction changes after dispatch has started?
- Why do effective dates matter when integrating change tasks into an agent environment?
- Which customer changes make useful interactive environment tasks?