Order processing: 400 orders a day, all entered manually
These days, an industrial supplier processes customer orders largely automatically. The processing time per order has fallen by 70 to 80 per cent.
- Mechanical engineering
- Inside sales
Key performance indicator
Processing time per order
11 min.<3min
Around 70 per cent of processes run without manual intervention
At a glance
Current situation
The internal sales team receives around 400 customer orders every day, mainly via email as PDF attachments, with a smaller proportion coming via customer portals and, in isolated cases, still by fax. The orders come from more than 1,500 active customers, each with their own form layout, product descriptions and reference numbers. Each order is read by a clerk, checked against the customer master data and the product database, and then manually created as an order in the ERP system. On average, this takes around 11 minutes per transaction.
- 400
- Orders per day
- 1.500+
- active customers
The challenge
Order processing was the bottleneck for the entire back-office team. During peak periods at the end of the month and quarter, orders would pile up overnight, and customers would not receive their order confirmations until the following day or later. Errors in item allocation – caused, for example, by customers’ own product names or outdated item numbers – led to incorrect deliveries, credit notes and rework in logistics and accounts. As the allocation process was largely based on experience, new staff could only work independently after several months, and the back-office team could only cope with growing order volumes by taking on additional staff.
Typical symptoms within the company
Why this had remained unresolved until now
Two previous attempts had already failed. After two years of maintenance, a traditional OCR and template system covered only the 30 most important customer layouts; any change to a customer’s form caused the recognition to fail. Whilst an RPA approach did automate data entry into the ERP system, it was unable to determine which item was being referred to. That was precisely where the real work lay: matching a free-text customer description to the correct item is a judgement call based on context, not a rule. With over 200 layouts and a high rate of exceptions, rule-based automation does not scale.
Our solution
Incoming orders are processed regardless of their format and without the need for a pre-defined template. Rather than searching for items at fixed coordinates, the document is interpreted based on its content: header data, items, quantities, prices and delivery terms are extracted as structured data, even if they appear in an unfamiliar tabular format or within running text. Item allocation is carried out via a semantic comparison against the item database, enriched with the respective customer’s historical orders. This enables the system to recognise the customer’s own terminology based on past ordering behaviour. Each item is assigned a confidence score. If this exceeds the threshold, the order is automatically created in the ERP system. If it falls below the threshold, the process proceeds to a review screen for the back-office team, displaying the highlighted item and the suggested allocation.
- Format-independent extraction of header and position data without the need for customer-specific templates
- Semantic product mapping against master data and the customer’s order history
- Confidence-based switch between background processing and manual checking
- Preview screen with pre-filled suggestions instead of a blank entry form
- Feedback: every correction made during the verification stage improves the matching process for this customer
From the inbox to an order in the ERP system
Incoming and identification
Orders from the mailbox, portal and scan are consolidated into a single inbox and classified as orders, enquiries or other.
Extraction
Header data and all line items are extracted in a structured format, regardless of the document’s layout and language.
Reconciliation and assessment
Customers, items and terms and conditions are matched against master data and historical records, and each line item is assigned a confidence score.
Classification
Transactions above the threshold are automatically created in the ERP system. All others are sent to the review screen.
Check by the back-office team
Only the highlighted items are checked and confirmed. The correction is incorporated into the reconciliation.
Measured over the first 6 months following roll-out
| Indicator | Value | Basis |
|---|---|---|
| Processing time per orderKey performance indicator | 11 mins → under 3 mins | Time recorded before/aftermeasured |
| Dark processing rate | around 70 per cent of processes without manual intervention | System measurementmeasured |
| Incorrect categorisation of items | around 60 to 70 per cent less | Random samplemeasured |
| Order confirmation on the date of receipt | from less than half to almost all transactions | System measurementmeasured |
| Induction of new staff | significantly shortened | User surveyCustomer details |
Cost-effectiveness
Integration and operation
Portability
This model applies wherever structured data needs to be extracted from documents without a fixed format and matched against an internal master data set, and where the matching process requires judgement rather than fixed rules. The decisive factor is not the industry, but the combination of high volume, high format variability and existing historical data as a source of context.
“We’re now checking for exceptions rather than typing everything out. The workload hasn’t decreased; it’s just different.”
Head of Internal Sales, Industrial Company
Similar situation in your own organisation? Let’s talk about it.
Briefly describe your process. We’ll get back to you with an initial assessment of whether the pattern can be replicated and what is realistically required.