Starting point
Krüger Metalworks (fictional) manufactures custom components for machine-building customers and takes in about 60 orders a week, by email and, more recently, also by WhatsApp for short-notice reorders. A staff member typed every order into the ERP manually: check article number, capture quantity, estimate delivery date based on capacity, send confirmation.
With follow-up questions (missing material, unclear quantity), the order often sat until the staff member found time to chase it themselves. Two to three times a month, a follow-up question simply got lost between emails, calls and the actual ERP upkeep, with the result that a customer only noticed their order hadn’t been processed by chasing it up themselves.
The setup (architecture as a list)
- Inbound channels: existing order inbox (email) and WhatsApp Business for reorders, both connected to an automation workflow.
- Matching: the agent reads article number, quantity, material and delivery date, and matches them against master data and stock levels in the ERP.
- Completeness check: if a required field is missing, the agent automatically creates a follow-up question to the customer and flags the order as “incomplete” instead of ignoring it.
- ERP write access: complete orders are automatically created as a job, including a confirmation email with a realistic delivery date.
- Log: every action (order created, follow-up sent, reply received) is traceable with a timestamp in the automation-software log.
A Tuesday with the system
7:50 am: an order arrives by email, article number, quantity and material complete. The agent matches it against the ERP, checks capacity and creates the job automatically. Confirmation out before the staff member has even had their first coffee.
9:15 am: a WhatsApp message orders “50 pieces like last time”, without an article number. The agent spots the gap and asks specifically for the article number, instead of guessing or letting the case sit.
2:00 pm: the staff member opens the console and sees at a glance two open follow-up questions, the rest already processed automatically, instead of digging through a full inbox.
Model calculation
| Assumption | Value |
|---|---|
| Orders/week (email + WhatsApp) | 60 |
| Avg. processing time manually (typing, ERP upkeep, follow-up if needed) | 14 min. |
| Avg. processing time with agent (only review & approval of special cases) | 3 min. |
| Time saved per order | 11 min. |
| Order-processing time saved/week | ≈ 660 min. → 11 hrs |
| Orders automatically created in the ERP | 60 of 60 → 100% |
| Before: forgotten follow-up questions/month | 2–3 |
| Now: forgotten follow-up questions (logged, none open without a flag) | 0 |
Limits & learnings
The agent doesn’t capture special commercial terms: volume discounts and special prices are still set by the staff member; the agent only handles what’s stored in master data. With the first WhatsApp messages using very sparse free text (“like last time”), the completeness-check rule needed tightening, because the agent initially interpreted too generously. And during a brief ERP maintenance window, the manual fallback kicked in as planned, because an automation project always needs a plan B, not a black box with no way out.