QUICK OVERVIEW
What to take away from the article
- The assistant should answer from the verified data of the specific company and order.
- The price is calculated by the price list and rules; AI prepares understandable text from them.
- Extend autonomy by message types according to the measured quality and severity of errors.
First, match the message to the client and the order
One mailbox often contains inquiries, invoices, complaint messages and communication with production. Before generating a response, the system must know which company the message belongs to and whether it is related to an existing order. The e-mail address match alone is not enough: one client can order several projects at the same time.
The recommended design combines job identifier, thread history and participant control. If he has more candidates, the assignment remains for approval. Attachments and message content are stored with origin and time to trace back why the assistant chose a particular response. The article on linking emails, orders and warehouse explains the practical basis.
For mailboxes on Websupport, the available connection method and authorization for a specific service must be verified. A connector intended for Gmail cannot automatically be considered a connector for another mailbox.
The delivery date must have a source and a degree of certainty
The supplier can write that he will hand over the shipment to the carrier on Tuesday. This is not yet the date of delivery to your client. The process model must distinguish between the production plan, the confirmed shipment, the delivery to you and the possible further shipment. The estimate is based on the current condition, usual transport on the given route and your handling reserve.
Fictitious example: the supplier confirms the shipment on Tuesday, the planned transport to you takes two working days, the inspection and repackaging takes one working day, and the subsequent delivery to the client takes one working day. Without holidays and with a follow-up pick-up, approximate delivery is on Monday. It is not a guaranteed date. When tracking is changed, the system must update the estimate.
It is enough to explain the status and approximate delivery window to the customer. Internal business information is not added automatically, but the system must retain it for its own review. If there is no confirmation of sending, the correct answer will admit that you are still verifying the date.
The price list decides the price, the AI decides the formulation
The model must not create a price based on the probability of the next word. The basis is the identified product or service, the valid version of the price list, the quantity, the client's discount and approved surcharges. The calculation should be repeatable: the same inputs will give the same result. AI can explain the options and prepare a nice offer, but it can't change the conditions silently.
Non-standard discounts, unclear shipping, missing dimensions or requests outside of the product range are to be approved. The proposal must show the used version of the price list and the reason for the exception. After approval, save both the exact document and the sent text. We discuss the bidding process in more detail in the guide to prices, scope and approval of bids.
Three levels of automation according to risk
Start with drafts that the worker will correct or confirm. Automation tools already support human approval before taking selected AI actions; n8n, for example, describes this with its Gmail node. This is a process control pattern, not a confirmation of compatibility with every email provider.
Design permissions by activities. Reading the status of the order, preparing a reply, changing the price and sending a message are four different powers. The text of the received e-mail must not grant itself a new authorization or override internal rules. The system must respect the company, the client and the role of the user even when searching for documents.
| Area | Postup | What to check |
|---|---|---|
| Proposal | Sorts the message and prepares a response with resources. | A person approves each submission. |
| Limited autonomy | Answers approved recurring questions from current data. | Exceptions are stopped, results are sampled. |
| Advanced process | Triggers exactly the allowed follow-up steps. | Limits, event auditing and immediate shutdown of automation. |
Learning from corrections must be measurable
A text correction is not automatically a new general instruction. A worker could change the tone for just one client or fix a one-time mistake. Record the intervention category: factual error, incorrect price, incorrect order, missing data or style. Repeated patterns then suggest modification of rules or documents.
First benchmark the change against an approved test set of messages, including unclear and conflicting cases. Track the proportion of responses without substantive correction, misassignment and serious errors. A single answer with the data of another client has a different weight than an inappropriate address. When the quality deteriorates, it must be possible to return the previous rules and limit autonomy.
How to find out if the assistant really saves
Measure the time from message receipt to actionable response and human time spent reviewing. Add rule maintenance, exception handling, and tooling costs. If preparation saves five minutes, but checking a new bug takes ten, the process isn't working well yet.
The pilot should have a narrow scope, for example questions about the production status of orders with complete data. Add price offers or complaints only after verification. We explain the difference between freed up capacity and realistically lower expenses in AI Automation Savings Model.
