A language model can produce a convincing answer without access to the actual state of a customer, inventory, contract or work order. A business user needs more than good text; they need an answer that distinguishes plan from execution, a valid document from an old version and a recommendation from an approved transaction. That context usually lives in ERP and connected systems.
This does not mean the model should receive unrestricted ERP access. A good design prepares bounded context for one task. A user permitted to see their customer portfolio should not gain access to the complete database through AI. An agent permitted to create a task does not need authority to change a financial transaction independently.
Level one is a sourced answer: AI finds and explains, while the user acts. Level two is a recommendation: the system prepares the next step and displays reasons, but a person confirms it. Level three is a bounded action: an agent executes a pre-approved step and stops when a condition is missing or an exception appears.
Users often need to see which data was used, which conditions were applied, what the system did not find and who confirmed an action. The OECD transparency principle emphasises context-appropriate information, including input sources and factors leading to a recommendation where feasible and useful. In a business process, this becomes a clear trace of sources, rules and accountability.
Source: OECD.AI: transparency and explainability principle — The principle is used as a public governance reference, not as a claim that ORKA solutions are certified against it.
ERP does not make AI trustworthy automatically. It provides a place to organise facts, permissions and consequences before an answer becomes an action.