Sample and Order Status Agents: Read-Only Logistics/ERP, No Invented ETAs

A foreign-trade sample and order-status Agent answers “where is my sample?” and “what is the order status?” by calling carrier/ERP read-only APIs and explaining status codes from a governed KB—or refusing and handing to ops on failure. It is not a soothing chatbot: invented ETAs create false expectations and complaints. Overview: /blog/what-is-foreign-trade-ai-agent. Sample-fee confirms stay human in the inquiry playbook: /blog/foreign-trade-inquiry-agent-playbook. Store logistics patterns: /blog/cross-border-ecommerce-agent—trade stresses sample IDs, PI/contract numbers, and ops handoff. GeonAI delivers acceptables; /agents are capability references only (not a public trial of 363+ presets). Email [email protected].

Why invented ETAs are forbidden

• Buyers plan customs/lines on fake arrival dates → disputes and claims
• “Should arrive soon” cannot be audited or reproduced in QA
• After tool timeout, model “completes” tracking → conflicts with carrier truth
• Wrong ETAs after sample fees paid destroy trust for the commercial order

Allow / deny (SOW-ready)

Write ACL and CRM binding: /blog/foreign-trade-agent-crm-follow-up. Contract boundary tables: /blog/custom-agent-scope-boundary-in-contract. Tool/MCP hard rules: /blog/agent-function-call-mcp-integration.

Recommended flow

1. Validate identity and IDs (order / tracking / sample request)—no key → refuse lookup and collect slots
2. Read-only carrier or ERP call; response includes resultCode, node times, raw status
3. Hit → templated latest official node + KB explanation (cited)
4. Miss / timeout / error code → fixed fail copy + hand to ops (optional CRM task)
5. Buyer pushes for arrival date → repeat latest node, no date fill; escalate to human

Fail-copy example

“No carrier receipt yet / system temporarily unavailable. We’ve routed this to our ops teammate for a reply within business hours. Please keep your reference number.”—no speculative dates or ports.

Minimum audit fields

• toolName, hashed shipment/order ID, resultCode, latency
• Handoff flag, thread ID, CRM lead ID
• Sampling: zero invented ETAs; fail path always uses refuse template

Boundaries vs inquiry / quotation Agents

• This Agent: status read-only—no numeric prices, no sample-fee confirms
• Inquiry Agent: first reply through sample progression; fees stay human (playbook)
• Quotation Agent: price-book guardrails and formal quote issue
• Split roles when permissions collide: /blog/multi-agent-vs-single-assistant

Quotation guardrails: /blog/foreign-trade-quotation-agent-guardrails. Multilingual replies still require read-only tracking: /blog/multilingual-b2b-trade-support-agent.

Anti-patterns (stop the project)

• Prompt says “give the buyer an approximate arrival date”
• Treat OCR of screenshots as official tracking without API validation
• Allow address/consignee edits without approval
• Promise “system found it” with no audit log

Pilot acceptance

• Invented ETAs / fake nodes = 0 (or scale-stop)
• On tool failure, 100% refuse-template use (sampled)
• Lookup success rate with valid IDs meets target; no “in transit” without shipment ID
• If writes are off, zero unauthorized writes; if on, human-confirm records exist

Pilot KPI framework: /blog/custom-agent-30-day-pilot-metrics.

How to brief GeonAI

Name ERP/OMS and carriers, whether read-only credentials exist, sample ID rules, and ops handoff channel. Email [email protected], /pricing, or Live chat. We deliver read-only guardrails—not “we can always give an arrival date.”