Support agents
that know the order.

They are built over the help centre. That makes them able to explain the returns policy and unable to answer “where is my order”, which is widely reported as the most common contact reason in retail. The customer gets a policy paragraph when they wanted a delivery date, and the deflection statistics look fine while satisfaction falls.

The fix is integration rather than a better model. The agent needs read access to the order management system, the carrier tracking feed, the returns system and stock availability — and, for real deflection, write access to a bounded set of actions: initiate a return, reissue a label, update a delivery address before dispatch, apply a goodwill credit within a defined limit.

That write access is where the engineering seriousness is. Every action needs a permission boundary, a value limit, an audit record and a rollback path. An agent that can issue refunds without a cap is not an efficiency; it is an incident waiting for a prompt-injection attempt in a customer message.

Handoff quality, more than deflection rate. A customer escalated to a human who then asks them to repeat everything has had a worse experience than if the agent had never existed. The transcript, the retrieved order state and the reason for escalation go to the agent, and that single detail moves satisfaction more than raw accuracy does.

Treat customer text as untrusted input. A message is data, not instruction. An agent with the ability to issue credits will eventually receive a message engineered to trigger one. Permission boundaries are enforced outside the model, in the tool layer, where a prompt cannot reach them.

Peak is the real test. Retail volume is spiky in a way that is entirely predictable — a promotional weekend, a delivery failure, a viral product. Capacity and graceful degradation under those conditions matter more than average-case latency, and are what a pilot in a quiet month will not tell you.