AI Order Entry: Orders from Email and PDF into the ERP

In B2B, orders rarely arrive the way the ERP needs them. They sit in the body of an email, in a PDF from the customer's purchasing system, on a scanned order form or in an Excel list. Someone in inside sales reads the order, looks up the customer number and the items and types the order in by hand. That is exactly where sales slows down: 39 percent of 200 surveyed wholesalers and manufacturers name the heavy manual effort in entering quotes and orders as one of the biggest brakes on their sales processes (ECC KÖLN). This article shows how AI-assisted order entry reads orders from email and PDF, matches them against customer and item master data and creates them in the ERP as orders awaiting approval, where fixed rules check and where a person keeps the final say. We describe the technical basis on our page on AI document processing.

Why order entry becomes the bottleneck

For the B2BEST Barometer, ECC KÖLN surveyed 200 wholesalers and manufacturers on sales and sales support in July 2026. Around four in ten companies (39 percent) spend more than 40 percent of their sales time on administrative and recurring tasks, and for 8 percent it is even more than 60 percent (ECC KÖLN).

The brake named most often is the heavy manual effort in entering quotes and orders: 39 percent of respondents count it among the biggest brakes on their sales processes (ECC KÖLN). Missing links between sales, online shop, ERP and CRM follow with 34 percent, complex internal coordination with 33 percent (ECC KÖLN). The two are connected: where the shop, the ERP and the inbox are not linked to each other, every order that does not come through the shop turns into manual work.

At the same time, customers expect a faster pace. 74 percent of the companies surveyed see a high risk of losing customers if enquiries are not answered on the same day (ECC KÖLN). An order that arrives at half past four and is typed in the next morning gets its order confirmation a day later. On top of that come typing errors that only surface during picking or invoicing: a mixed-up customer item number, cartons instead of pieces, a price from the old list.

Structured, semi-structured, free: where orders come from

Not every order needs AI. If you exchange orders with a customer as EDI messages, they already arrive in an agreed, machine-readable format; how that works is covered in our article on EDI integration. Orders from the B2B shop are also created in structured form, for example through a quick order for regular customers. For everything else, the route into the ERP usually still runs through the keyboard today.

This article deals with the bottom four rows. They do not disappear even when a shop exists: many buyers order from within their own system, and its purchase order arrives as a PDF. Digital self-service is also often still being built up: a third of the wholesalers and manufacturers surveyed offer comprehensive self-service, while 56 percent only have basic functions such as simple ordering options (ECC KÖLN). If you want to handle enquiries rather than orders digitally, the article on the RFQ workflow in the B2B shop describes the route.

How AI reads an order

Technically, AI-assisted order entry is not a single function but a chain of clearly separated steps, each with its own check, so that an error does not pass on unnoticed.

1. Collect the mailbox: A service fetches the order mailbox, stores the email and attachments unchanged and assigns a case number. Duplicate submissions are detected.
2. Classify: Is it an order, a change, a cancellation, an enquiry, a complaint or a delivery notice? Only orders and changes continue; everything else goes to the team responsible.
3. Extract: Header data such as sender, the customer's order number, order date, delivery address and requested date, plus the lines with customer item number, description, quantity, unit and, if stated, price. PDFs with a text layer are read directly, scans and faxes via text recognition.
4. Match: Customer and items are resolved against the master data in the ERP, units are converted and prices compared.
5. Check: Fixed rules and the confidence per field decide whether the order goes to quick confirmation or to detailed review.
6. Hand over: The order is created in the ERP, initially in a status that requires approval.

Extraction uses language models that process text and table layout together. What matters is what they return: not free text, but a fixed data format in which each field carries its value, its location in the document and an estimate of how certain the assignment is. Ordinary program code then checks this format: is the quantity a number, does the unit exist, does the date make sense? What the model delivers is a proposal, not a booking.

Text analysis is by no means a niche any more. 26 percent of companies in Germany with at least ten employees used AI in 2025 (Destatis). Of these, 52 percent used text mining, and 27 percent used AI to automate workflows or decisions (Destatis). The building blocks that read an order are therefore among the most widespread AI applications; the effort lies in embedding them in your processes. How text analysis assigns products to the right categories in the shop automatically is shown on our page on AI category automation.

Matching against customer and item master data

Extraction is the easier part. The harder question is which customer and which item are meant. An order names the customer by company name and sender address, rarely by your customer number. And it names items in the customer's language: with their own material number, a shortened description or the manufacturer's part number. Matching translates these details into your master data.

The mapping table is the real asset of the system. With every confirmed order it knows more customer item numbers, and the share of lines that are mapped without a query rises. What matters is where this table lives: in the ERP or in a separate database, not in the memory of a model. Then every mapping stays traceable and correctable.

Plausibility checks: rules before every approval

Confidence per field is not enough on its own. A model can be very sure about an order that it read correctly but that makes no business sense: ten times the usual quantity, delivery to an unfamiliar address, a price from last year. That is why a set of fixed rules runs after extraction, rules your sales team knows and that can be traced without AI. Every rule that fails sends the order to review and states the reason.

• Duplicate order: the same customer order number or the same lines within a short interval
• Quantity outside this customer's usual range for this item
• Minimum quantity or pack size not met
• Price in the order differs from the stored customer price
• Requested date is earlier than the possible delivery time
• Delivery address is new or differs from the stored addresses
• Credit limit exceeded or customer blocked, see managing credit limits in B2B
• Item discontinued, blocked or not released for this customer

The rules also have a legal side. In many cases, the customer's order is the offer and your order confirmation is the acceptance. An offer made to a person who is absent can only be accepted until the time at which the offeror may expect to receive the answer under ordinary circumstances (Section 147 (2) BGB), and a late acceptance is considered a new offer (Section 150 (1) BGB). A confirmation that departs from the offer is treated by the law as a rejection:

An acceptance with expansions, restrictions or other alterations is deemed to be a rejection combined with a new offer.

Section 150 (2) BGB, English translation on gesetze-im-internet.de

A difference that is silently corrected during entry therefore changes the contract you offer the customer. That is why every difference belongs visibly in the review screen and, if it remains, explicitly in the order confirmation. The legal classification in an individual case, for example after a quotation you issued earlier, is a matter for your legal advisers.

Human approval: who decides what

The review screen is where inside sales works: the original on the left, the extracted fields on the right, uncertain or conspicuous fields flagged and given a reason. Confirming takes one click, corrections are logged and flow back into the mapping table. Responsibility stays with a person without every line having to be typed again.

You decide where the boundaries between the levels lie, based on measurement rather than gut feeling. In the pilot, the system counts for each field how often the extracted value was right without correction. Only when a customer, a field and a group of items have been stable for weeks is it worth reducing the review there to spot checks. Amounts and delivery addresses usually stay in review. How thresholds and approvals work in automated decisions in general is described on our page on automated decision-making.

Approval is not a brake here: what gets checked is a prepared order, and the order confirmation can go out on the same day. 78 percent of the 200 wholesalers and manufacturers surveyed in March 2026 see great potential in AI for handling customer enquiries faster and with higher quality (ECC KÖLN), and 77 percent of those surveyed in July believe that AI has significantly raised expectations of fast response times (ECC KÖLN). By contrast, there is usually no legal obligation to confirm receipt of an order received by email electronically and without undue delay: the corresponding duties in electronic commerce do not apply if the contract is concluded exclusively by individual communication (Section 312i (2) BGB). The time pressure comes from customer expectations, not from the law.

Handover to the ERP: create instead of type

The end result is an order in the ERP, not a PDF with highlights. That requires an interface that can create orders with header and line data, and a clear rule on the status in which they arrive. A dedicated status such as “awaiting approval” or a pre-entry area from which inside sales releases the order has proven useful. After that, everything runs as usual: availability check, order confirmation from the ERP, picking and invoicing.

• SAP Business One: create sales orders via the Service Layer, details on SAP Business One integration
• Microsoft Dynamics 365 Business Central and Dynamics NAV: sales orders via the APIs or web services, see Microsoft Dynamics integration. What the end of support for Dynamics NAV 2017 means for existing integrations is covered in Dynamics NAV 2017: end of support and shop integration.
• JTL-Wawi: create orders via the interfaces of the inventory management system
• Other or in-house systems: via an individually developed connection

The interface writes back: order number, status and differences are stored with the case, and the original order is linked to the order in the ERP. The bottleneck is rarely the model but the integration: 44 percent of the wholesalers and manufacturers surveyed name integration into existing systems as one of the biggest hurdles to using AI (ECC KÖLN). An overview of the systems we connect is on our integrations page.

Security: when the order itself is the attack

An order in the mailbox is external content. If you have a language model read it, you have to expect that someone will try to steer the model through exactly this content. The OWASP project's list of risks for applications built on language models lists prompt injection under the identifier LLM01:2025 and explicitly describes the indirect variant, in which a model processes input from external sources such as websites or files (OWASP). Such instructions do not have to be visible to humans; it is enough for the model to process the content (OWASP). White text on a white background in a PDF is an obvious example.

Independently of AI, there are fraud attempts with orders that, in the name of known customers, order goods to a new delivery address. The rule “new delivery address goes to review” protects here regardless of how well the model reads. How AI features can be protected against manipulated input in general is described in Prompt injection: securing AI features in your shop.

Data protection, operation and retention

Orders contain personal data: names and phone numbers of buyers, delivery addresses. If processing takes place at a service provider, a data processing agreement is needed that sets out, among other things, the subject matter, duration, nature and purpose of the processing (Art. 28 (3) GDPR). If you do not want the documents to leave the company at all, you run the models on your own infrastructure; what that means technically is described in AI in your shop without data leaving the house. The barometer shows that the question carries weight: 37 percent of the wholesalers and manufacturers surveyed name data protection and security risks as a hurdle to using AI (ECC KÖLN).

Then there is retention. An order received by email is usually an incoming commercial letter. Merchants keep incoming commercial letters for six years (Section 257 (4) HGB), and for tax purposes a six-year period also applies to incoming commercial or business letters (Section 147 (3) AO). If documents are kept electronically, the reproduction must match the incoming commercial letter visually when it is made readable (Section 257 (3) HGB). For order entry, this means that the original stays unchanged alongside the order, and the extracted values and every correction are logged in addition. The tax and legal assessment in an individual case lies with your tax adviser.

Rollout in four steps

1. Count and sort orders: Analyze a few weeks of incoming mail: channels, formats, customers and the share of orders that needed a query. This gives the starting point, often the PDF orders from the customers with the highest volume.
2. Define rules and responsibilities: Which checks apply, which thresholds per field, who approves and what happens to orders that arrive outside working hours?
3. Pilot alongside the existing route: The AI prepares orders while inside sales continues to enter them as usual at first, and the comparison is made per field.
4. Operation and expansion: Further customers and formats are added step by step, and where the measurements support it, spot checks replace full review.

Many companies are at exactly this threshold: 39 percent of the wholesalers and manufacturers surveyed in March 2026 already use AI to automate internal workflows (ECC KÖLN). But AI only becomes a tool that inside sales can rely on through measurement and a clean connection to the ERP.

Metrics for measuring the benefit

A few metrics show whether the change pays off. It is important to collect them for the existing route before the pilot, otherwise there is nothing to compare against.

The effect shows up mainly in sales: time that currently goes into typing becomes available for advice and for enquiries that require more than a confirmation. How order entry, shop and ERP work together in B2B sales is described on our page on B2B e-commerce.

How we implement AI order entry

We develop order entry individually for your order channels and your ERP. It is based on our AI document processing: collecting the mailbox, classifying and reading documents, showing the confidence per field, presenting uncertain cases in a review screen and handing the record over to the target system.

Confirm faster without giving up control

Orders by email and PDF will remain part of everyday B2B business for the foreseeable future, even alongside EDI and the shop. AI-assisted order entry shifts the work they cause from typing to checking: it reads, matches and prepares, fixed rules catch what does not fit commercially, and inside sales decides. The order confirmation can go out sooner, and the decision stays in-house.