Orders, delivery notes and forms into your systems without retyping

Your clerks stop copying order numbers, quantities and dates from PDFs into SAP and spend their time on the exceptions: the wrong price, the unknown material number, the customer who needs a call. Computerwoche described the typical starting point at Zentis in 2023: many orders arrive as email with a PDF attached, so they cannot be processed automatically in SAP and have to be entered again by hand. The trade magazine logistik-heute, reporting on the logistics software vendor Nemox, summed up the everyday situation in 2026: a clerk today works in five programs at once.

Text is where German companies use AI most. Among companies with 10 or more employees that used AI, Destatis counted text mining at 48% and natural language generation at 34% in 2024. Sector surveys put extraction near the top: in a 2025 survey of 35 insurers in several countries by Sollers Consulting, 69% used AI for data extraction, and in a 2026 Descartes benchmark of 300 European shippers and logistics providers, a vendor study, 41% used AI for automated data entry and processing unstructured data.

• Fewer hours of typing per order, delivery note or claim. We measure hours per document before and after, on your own volume.
• Orders are in SAP as soon as they arrive, without waiting for someone to get to the inbox, so confirmations go out sooner.
• Fewer typing errors in quantities, prices and dates that later turn into wrong deliveries and credit notes.
• Peaks at season start or month end are handled without extra temporary staff typing.
• Experienced clerks spend their time on customers and exceptions, which matters when the skills shortage makes them hard to replace.

The system reads the documents that arrive by email, scan or upload, extracts the fields your process needs and prepares the booking or the case for the next step.

Two measured examples from other companies, reported by the press and by vendors. Cloudsail was involved in neither. Computerwoche reported in 2023 that Zentis in Aachen reads orders that arrive as email with PDF using the vendor Workist: the share of documents processed automatically rose from about 60% after training to about 80%, with field recognition over 90%. In 2026 the vendor Nemox reported in the trade magazine logistik-heute that the Frankfurt air freight forwarder Cargo Movers reconciled more than 28000 freight invoice lines against its transport management system, and 97% needed no correction of the extracted values.

• Customer orders by email or PDF: customer, material numbers, quantities, prices and delivery dates, prepared as SAP sales orders.
• Delivery notes: matched to the purchase order and goods receipt, with differences in quantity or item flagged.
• Supplier order confirmations: confirmed dates and prices compared with your order, with deviations sent to purchasing.
• Claims and complaints: customer or policy number, description, amounts and attachments sorted, with duplicates detected.
• Applications and forms, such as customer, service, permit or subsidy applications, checked for completeness before a person decides.
• Scanned forms such as meter reading cards, SEPA mandates and bank detail changes. The trade press (ZfK, 2026) reported that Stadtwerke Pforzheim is working on exactly these.
• All other incoming mail: sorted by type and routed to the right team, with the main fields already filled in.

The system can check incoming contracts and supplier terms against your company's playbook and mark every clause that deviates, with the clause and your rule side by side. Your lawyer or contract manager decides. A playbook here means your written rules: the liability caps you accept, payment terms, penalties, governing law, termination and data protection clauses, with fallback positions. The check finds the matching clause for each rule and marks it as matching, deviating or missing, and it can draft a reply in your standard wording for the reviewer to edit.

Interest is ahead of use. In Bitkom's 2026 survey, 8% of the companies that use AI use it in the legal or tax department. In Noerr's 2024 survey of in-house legal departments, about 25% already used AI tools, only 15% had AI guidelines and 42% saw AI as a security risk.

Contracts carry trade secrets, so this check runs on your own servers or in your own cloud account. Law firms usually turn to legal AI products such as Noxtua or Harvey. A custom check pays off for in-house teams with their own playbook, contracts that must stay in the company, and a DMS such as d.velop or ELO it has to work with. What comes out is a prepared review for your lawyer, who stays responsible for it. It is not legal advice.

People check what the system is unsure about and decide everything that commits the company: prices, delivery promises, payments, claims and contract terms.

• Every field carries a confidence value. Below the agreed threshold, the document goes to a review queue with the field highlighted on the original.
• Checks against master data stop silent errors: an unknown material number, a price off the price list or a quantity far outside the customer's usual order goes to a person.
• New customers, new suppliers and documents of an unknown type always go to a person.
• At the start everything is reviewed. Document types are released for straight-through processing one by one, once their measured accuracy meets the agreed threshold.
• Credit decisions, claim settlements and payments stay with people. Assessing the creditworthiness of natural persons and risk assessment and pricing in life and health insurance are high-risk uses under Annex III of the AI Act, and a company that uses AI for them takes on the high-risk duties, including human oversight, which apply from 2 December 2027 under the current timeline. Reading data from such an application is extraction. We do not automate the decision itself.
• Job applications are not in scope: screening and ranking applicants is also high-risk under the AI Act, and we do not automate that decision.

We connect the inbox or scanner, read each document with a method suited to its type, validate the fields against your systems and hand the result over through the interfaces your ERP already offers. Documents and extracted data stay on your own servers or in your own cloud account.

• Intake: shared mailboxes in Outlook or Exchange, scanner output, upload folders, a DMS such as d.velop or ELO, and EDI where partners already send structured data.
• Reading: text recognition for scans, plus a language model that understands layouts it has not seen before, so a new customer's order form does not need a new template.
• Validation: customer, material and supplier numbers looked up in SAP or your ERP, prices compared with price lists and contracts, totals and dates checked.
• Review queue: one screen with the original document next to the extracted fields. Every correction is stored and becomes a new test case.
• Handover: SAP sales orders, order confirmations or goods receipts go in through standard SAP interfaces. Accounting data goes into DATEV by the route your accounting or tax adviser already uses, and other ERPs are fed through their interfaces or import files. The original stays linked in your DMS.
• Operation: monitoring of volumes, queue length and correction rates, with alerts when a customer changes their form and accuracy drops.

If you run SAP S/4HANA Cloud, check the built-in feature first. According to the SAP partner PIKON (2025), S/4HANA Cloud already creates sales order requests from PDF or image, suggests field values and simulates price or item deviations before conversion. For standard orders into one current SAP system, that may be enough.

Specialised document processing products, often called IDP for intelligent document processing, cover common document types such as invoices well. A custom build is justified when one of these applies, and we tell you in the first call if buying is the better answer:

• Several ERPs, an older SAP release, or SAP mixed with other systems.
• Document types no product knows: your own forms, technical delivery notes, claims with mixed attachments.
• Checks that need your rules: contract prices, customer-specific agreements, plausibility against order history.
• Documents that must stay on your own servers.
• One inbox with many document types that need sorting before they can be read.

Before go-live we measure, per field and per document type, how often the system extracts the right value, on a sample of your own documents. Vendor figures from other companies tell you little about your forms, so the decision rests on your numbers.

After go-live we report corrections per field and document type, so a drop shows up when a customer or supplier changes their layout. Where you want a business case, we measure hours per document before and after.

• A test set from your real documents, including poor scans, handwritten notes and unusual layouts, with the correct values taken from what your staff entered.
• Field-level accuracy: order number, material number, quantity, price and date measured separately, because a wrong quantity costs more than a wrong reference.
• The share of documents that need no correction at all, per document type.
• Validation tests: documents with deliberate errors, such as an unknown material or a wrong price, must land in the review queue.
• Shadow mode before go-live: the system runs next to your clerks, who keep entering data as usual, and both results are compared.
• Thresholds agreed in writing before the acceptance test, and the same tests after every change to model, prompts or rules.

Orders, claims and forms often contain personal data, so we document what is stored, where it is processed, how long document images and extracted data are kept, and who can see them. You decide, advised by your data protection officer and lawyers. What we give you is a technical description and not legal advice. Every extracted value keeps a link to the original document and to the person who approved it, so auditors and your tax adviser can trace each booking.

E-invoicing changes which invoices still need reading: XRechnung and ZUGFeRD invoices arrive as structured data, so there is less and less reading to do on domestic invoices. Our finance and accounting page covers the e-invoicing dates and the work that remains there, checking invoices against orders and goods receipts. Document processing keeps all the documents that are not invoices: orders, delivery notes, confirmations, claims, applications, and invoices from abroad.

If the review queue records who corrected what, it can show individual performance, which brings in the works council under Section 87(1) no. 6 of the Works Constitution Act (BetrVG). We log per document and team with pseudonymised user IDs and agree any reports with the works council before go-live.

Start with the one document type that costs the most typing and has clear rules, usually customer orders or supplier confirmations. We agree a fixed price for each step in writing after the first call.

• Count documents per type and the hours spent on them today.
• Collect a sample of real documents with the values your staff entered for them.
• Check whether SAP's built-in feature or a product already covers that type.
• Agree fields, thresholds and review rules, then build, test in shadow mode and go live.
• Add the next document type with its own tests.