Aftersales is full of context: policy, vehicle history, parts status, customer promises, technical evidence, commercial exposure and role-specific authority. A generic chatbot can summarise text; a useful assistant has to respect the operating model.
Start with evidence boundaries
If the data does not support a conclusion, the system should say so. Missing evidence should reduce scope, not be silently converted into certainty.
Keep deterministic rules where they belong
Some controls should not be delegated to generative reasoning: entitlement, payment, safety gates, calculation rules, policy thresholds and irreversible actions need explicit logic and authority.
Make the role visible
A parts manager, warranty administrator and dealer principal need different levels of detail and different decisions. Role-aware intelligence should translate the same evidence into the right operational context without changing the underlying facts.
Human authority is a feature
AI can propose, summarise, challenge and coach. The accountable person should still own the decision — especially where customer, financial or safety consequences are material.
The test
If the assistant cannot explain what evidence it used, what it does not know and who must decide next, it is not ready for operational authority.
The assistant must know when to stop
Imagine a parts assistant identifying apparent excess stock but lacking purchase history and supplier return terms. It can explain the arithmetic and suggest an investigation. It cannot responsibly turn that finding into an instruction to dispose of the stock. Calculation confidence and action confidence are different, and the interface should make that difference visible.
Test the assistant with incomplete evidence, conflicting records and requests outside the user's authority. A useful response names the missing fact, narrows the recommendation and identifies who can decide next. It should not fill the gap with plausible language. Retain deterministic controls for permission and financial calculations, with AI contributing interpretation within those boundaries.
Jano remains a developing capability. These principles describe the intended operating standard, not proof that every integration is live. Evaluate the system through the decisions it helps people make and the unsupported decisions it prevents. The commercial goal is less avoidable checking and clearer accountability, with confidence grounded in evidence rather than fluent presentation.