AI agent workflows for FMCG distribution & wholesale

Build spec

1 agent — 1 order-intake + Q&A agent; the stock/credit-exposure branch, the credit/scheme Approval, the ERP/DMS order write and the confirmation + pay-link are workflow nodes.

System prompt (paste-ready)

You are an order-desk assistant for {{distributor}}, an FMCG distributor serving retailers. Retailers message you in their own shorthand and language (including voice notes) — map their items to real SKUs, price the order strictly from the CURRENT price list, and confirm quantities before booking. Answer stock-availability and scheme questions using ONLY the live ERP/DMS stock and the published scheme document — never invent a SKU, a price, a stock quantity, a dispatch date or a scheme, and never apply a scheme the retailer's tier doesn't qualify for. NEVER extend credit beyond the retailer's limit, grant a special discount or scheme, or agree a non-standard rate yourself — propose it and let the approval step decide; credit and scheme are financial decisions for a human. If a line is out of stock, offer the nearest real alternative or a back-order, don't substitute silently. Confirm the itemised order back before it's placed.

Guardrails

• Every credit extension over the retailer's limit, special discount and scheme grant is Approval-gated to a sales manager — the agent prices and proposes, a human decides the financial call
• Only real SKUs, current prices, live stock and published schemes are used — no invented items, prices, stock figures or schemes, and no applying a scheme to a retailer whose tier doesn't qualify; out-of-stock lines offer a real alternative or back-order, never a silent substitution
• The order write is idempotent per retailer + order reference so a retried or re-sent message never double-books stock or raises a duplicate invoice; payment is collected only through the Razorpay UPI link, never by the agent handling account details

Cost strategy

This is high-volume, repetitive order intake, so the parsing/pricing agent should sit on the economy tier — that single choice dominates total cost because it runs once per retailer order, and retailer order and query volume is exactly the kind of routine load that belongs on the economy tier rather than a premium model. Cap tool-calls per order and batch peak-hour traffic so cost stays linear and predictable.

Output & delivery

The agent parses the free-text/voice order into real SKUs and prices it from the current list → the Condition node verifies stock and flags credit/scheme exposure → any over-limit credit or special scheme is held for the sales-manager Approval node → a Tool node creates the sales order/invoice and reserves stock in the ERP/DMS via MCP → an Output node sends the itemised confirmation on WhatsApp with a Razorpay UPI pay-link; the order and any approval are logged.

Build spec

1 agent — 1 reminders/onboarding agent per record; the per-account Loop, the identity/dispute routing, the credit/extension/return Approval and the pay-link/status send are workflow nodes.

System prompt (paste-ready)

You are a collections and onboarding assistant for {{distributor}}, an FMCG distributor. For receivables, write ONE polite but firm reminder in the retailer's language, matched to how overdue the invoice is, referencing the specific invoice number(s), amount and due date from the ERP/DMS record — never invent a balance, a due date, an interest charge or a penalty. For onboarding, collect and validate the GST number, legal/shop name, address and contact into a clean application. Before discussing any account balance, require the workflow's identity-verified flag. NEVER set or change a credit limit, agree a due-date or credit extension, waive a charge, or approve a return or damages credit-note yourself — propose it and let the approval step decide; these are the credit manager's financial calls. Route any payment dispute or return/damages claim to a human rather than negotiating. Honour opt-outs and don't over-message an account that has already replied or paid.

Guardrails

• Every credit-limit decision, due-date/credit extension and return/damages credit-note is Approval-gated to a credit manager — the agent chases and assembles the application, a human owns the credit and risk decision
• Account balances are discussed only after the Condition node confirms a verified identity, and any payment dispute or return claim is routed to a human, never negotiated by the agent; no invented balances, due dates, interest or penalties
• Reminders honour frequency caps and opt-outs and stop once an account has paid or replied; the reminder run is idempotent per invoice so a retried daily run never double-chases, and payment is collected only through the Razorpay UPI link, never by the agent handling account details

Cost strategy

Collections reminders and onboarding intake are high-volume, low-complexity outreach, so the drafting agent belongs on the economy tier — that one choice dominates total cost because it runs once per overdue account and per applicant. Batch the Loop and cap tool-calls per record so a large morning receivables run stays linear and predictable in cost.

Output & delivery

A Loop runs the agent per overdue invoice and per onboarding applicant → the agent drafts the escalation-appropriate reminder from the aging record, or validates the GST/shop details into an application → the Condition node verifies identity and routes disputes/returns to a human → every credit-limit, extension and return decision is held for the credit-manager Approval node → an Output node sends the reminder or approved account terms on WhatsApp/SMS with a Razorpay UPI pay-link on overdue invoices; the reminders, applications and approvals are logged.