Key terms explained
An integration exchanges data between systems. RPA (Robotic Process Automation) can perform actions through existing screens. When processing orders, you still need to check that the order has actually been recorded.
There is more to the problem than data entry
Imagine a wholesaler receiving orders in a shared inbox. One customer sends a PDF; another puts the order in the email itself. An administrator looks up the customer account, matches product descriptions to internal codes and enters the order.
The steps look predictable until a customer uses an old product code or orders twenty items when the product is sold by the box. At that point, better extraction is not the whole answer. Someone needs to establish what the customer means.
Start with a limited scope, such as repeat orders from existing customers for known products. Urgent orders, amendments and new customers can stay with a person for now. This makes it easier to see which work is genuinely repeatable.
Agree what makes an order ready to process
Work with the order administration team to define the required information and checks. In this example:
- The customer and delivery address are known and appropriate for the order.
- A customer reference is available, or an agreed alternative is used.
- The product code, quantity and ordering unit are unambiguous.
- The requested delivery date and any special arrangements are clear.
- It is clear who can approve pricing, delivery or other exceptions.
A populated field is not necessarily a correct one. Check, for example, whether the product code exists and whether the quantity uses the agreed unit. Missing or unreadable information should not become a plausible guess.
Also agree how an authorised sender is identified. A familiar name at the top of an email is not, on its own, an adequate check.
What if the email and attachment disagree?
A fictional example: the email says ten boxes, while the attachment says ten items. Show both proposed values alongside their source. Do not automatically pick whichever value is easier to process. A person checks the intended quantity and unit with the customer.
Record where each field comes from: the message, an attachment or existing customer and product records. Agree how to link a later email requesting a change to the original order. A new message with the same customer name is not necessarily a new order. A missing delivery date should remain visibly missing; do not let AI invent an agreed date.
Design human review of an AI proposal or discuss which order step might benefit from AI.
One possible route from inbox to order system
The sequence below is a design example. Agree with the process owner which steps may run automatically.
Receive
Register the request and its source. Apply the agreed sender and attachment checks before processing the information.
Extract
Read the required fields from the message or document. Keep the connection to the source so a reviewer can see where a value came from.
Validate
Compare the fields with customer and product records and the agreed order rules. Set unclear cases aside.
Approve
Only pass orders that meet the agreed conditions. Where manual approval is required, the process waits for it.
Transfer
Pass the approved information to the order system through a suitable integration, import or other execution method.
Confirm
Check that the order has actually been recorded and retain its reference. Acknowledging an email is not the same as confirming a recorded order.
When the route changes
- Unclear information: clarify before approval.
- Repeated order: check whether it is a repeat or an amendment.
- No confirmation: investigate before submitting again.
An illustrative process. Exceptions return to a person before processing continues.
Three exceptions to decide on before you start
The messages and status notices below are fictional. They show which decision is still needed before an order can proceed. These are not results from a customer trial.
On a narrow screen, scroll the table sideways. With a keyboard, focus the table and use the arrow keys.
| Fictional example | Check and responsible role | What to avoid |
|---|---|---|
| Email: “Please send 20 of item A17.” It is unclear whether this means units or boxes. | Order administration asks for the unit and records the decision. The order remains on hold. | Guessing an unclear quantity and submitting the order anyway. |
| Email: “Sending order REF-042 again.” | Order administration compares it with the existing request: the same order, an amendment or a new order? The reference helps with the check; it does not prove uniqueness. | Treating every new email as a new order. |
| Status notice: “Transfer sent; confirmation missing.” | The designated administrator and order administration check whether the order was recorded. Resolve the uncertain status before deciding to resubmit. | Blindly submitting it again and potentially creating a second order. |
Do not let the exception queue become another unattended inbox. Decide who takes ownership, what information they need and when to escalate. A correction should be traceable without reconstructing the entire process.
Choose the technology once the process is clear
First check whether the existing order system offers a suitable import or integration. If it does not, interacting through its user interface may be an option to investigate. Availability, access permissions and ongoing maintenance all affect that decision.
AI is optional. Consistent input may be handled with rules. For varied documents, you can investigate whether AI helps prepare the information, provided uncertain results remain visible and cannot independently create unwanted orders.
Also establish what data an external service would receive and who could access it. Do not use real order documents in a trial environment before agreeing how data will be used and protected.
Clearer input may be the better first step
If missing information causes most of the rework, an agreed order format, an existing customer portal or a clear work instruction may already help. Automation does not have to be the first investment.
A small trial is most useful when you know which part you want to improve. Compare not just entry time, but also the effort spent checking orders, resolving exceptions and maintaining the solution.
What should a trial help you assess?
Use examples that represent the actual work, including exceptions. Agree in advance when to proceed and when to revise the approach.
- How many requests can follow the agreed route?
- How much manual checking and rework remains?
- Are repeated and amended orders handled correctly?
- Can you tell whether each request is waiting, rejected or recorded?
- Can the responsible person understand and resolve failures?
Extracting a document successfully does not prove that the order process works. Assess the outcome in the order system and keep uncertainties visible. This example does not establish a general savings percentage.
Five questions about your own order process
- Which types of orders recur regularly?
- What information does someone currently look up, correct or request?
- Which exceptions genuinely require a decision?
- How do you currently know that an order was recorded correctly and only once?
- Who owns the process when something goes wrong?
Does this sound like your order process?
Describe where orders get held up or where information is entered again. We can discuss which step is worth investigating. You do not need to have chosen a solution first.
Discuss your order processA defined trial: from request to confirmed test order
Start with repeat orders from existing customers for known products, one intake channel and one agreed ordering unit per product. New customers, pricing agreements, urgent orders and changes remain outside the automated route in this trial. Use fictional data in an agreed test environment; the trial creates no real orders.
Define input and output beforehand
- Input: a fictional email with customer reference, product, quantity, unit and requested date, plus approved test master data.
- Intermediate result: proposed fields linked to the source, a check outcome and a specified reason when the request is waiting.
- Final result: a confirmed test order with a system reference, or an explained waiting or rejection status. A sent message alone is not a final result.
On a narrow screen, focus the table and use the arrow keys to scroll.
| Test case | Expected outcome | Evidence and responsible role |
|---|---|---|
| O-01 · known customer, product A17, two boxes. | All agreed checks pass. One test order is created only after the required approval. | Order administration compares source, customer, product, two boxes and the reference in the test order system. |
| O-02 · “20 of A17”, with no unit. | The request waits; the process does not choose pieces or boxes itself. | Order administration sees the open question and records clarification before approval. |
| O-03 · the same message arrives again. | Investigate repetition versus change; a new message does not independently create another order. | Order administration compares the original request and existing order status; the reference alone does not decide. |
| O-04 · the destination processed it, but confirmation is missing. | The outcome remains unknown until investigated; no blind second attempt. | The administrator and order administration investigate destination status and record a recovery decision. |
Fictional record trail: source, decision and destination stay connected
Request O-01 names customer K-104 and two boxes of A17. Authorised employees can trace the source. The reviewer confirms that the customer, product and unit match the test master data and approves the test request under the agreed role. The test environment returns order T-501.
The final check compares the recorded lines with the approved request. Only then does O-01 become “recorded”. Missing confirmation puts the request into an investigation status with an owner. A correction retains its link to the first request; extracting the email again is not evidence of recovery.
Assess the complete route, not only extraction
- Have order administration define the expected outcome for each case beforehand and assess the evidence afterwards.
- Measure searching, review, entry and recovery within the same boundaries as the current practice; record waiting separately.
- Define which error immediately stops the trial, who follows up open cases and which observations are needed to proceed.
Download the blank order trial log (CSV)
The log contains suggested fictional cases; result fields are blank. Record references and observations, not real customer or supplier details. Downloading does not send data.
Define scope, quality and stopping criteria in the pilot checklist
Start with the process canvas if the handover is still unclear
The cases and record trail are our own fictional examples. The OWASP Input Validation Cheat Sheet describes checking the form and meaning of input (consulted 14 October 2026). This supports the validation principles, not the availability of a specific order integration.
Adapt this example in the process canvas
Start with fictional scope, systems, exceptions and open questions for this process. You review and confirm before the example replaces canvas answers. Adapt it with the process owner and save your own version locally.
