What German Suppliers Should Ask About AI Order Intake

Sales order intake in German industry runs on email: orders arrive as PDFs, Excel files and free-text messages, often with customer-specific article numbers and units. Many "AI" tools only extract fields and hand the work back to your team. Use these questions to tell autonomous processing apart from extraction.

Each question has two answers. One should reassure you, the other should worry you, and vendors rarely notice which one they just gave. Ask them cold and write down the exact wording. Sales order automation projects fail on the things that sound like details in a first call.

1. Does it read orders in the formats and languages you actually receive?

Ask for a live test on your own messy examples: German and English, PDFs, spreadsheets, free text, even handwritten notes. A capable system reads them without a template per layout. GeneralMind's InboxIQ classifies each incoming email and routes it to the right workflow (order, modification, invoice, claim) regardless of format or language.

"We support PDF" means nothing on its own, so ask which PDFs. A native export from a customer's ERP is easy. The case that decides whether a tool survives your inbox is the scan of a fax: handwritten delivery week in the margin, stamp across the article table. Ask what happens when the order is in the email body and the attachment is last quarter's price list, or when the subject says Bestellung and the real instruction is a two-line change three replies down.

Language support is not translation. German order mail carries abbreviations no dictionary resolves: Stk., VE, Pal., ab Werk, Liefertermin KW 34. The same buyer writes them differently in March and in September. If you run a mailbox per country, ask to see two languages inside one run. The answer to walk away from: send us fifty examples per customer and we will train on those. That is a template library with a friendlier name, and it bills you again whenever a customer changes system.

2. How does it match to your master data?

Extraction is easy; correct matching is hard. Ask how the tool maps a supplier's or customer's wording to the exact ERP entity, SKU and unit. GeneralMind uses context-based fuzzy matching, reconciling name, postcode and delivery address to identify the right entity and converting units (e.g. pallets → cases) before booking.

Name the mess out loud in the meeting. Your customer orders 04-7712 because that is what their system calls it; your material master knows it as 1099847-B, and the bridge between them is a spreadsheet one person maintains. The next customer sends no number at all and orders by description: 2 mm Blech, verzinkt, 1000 x 2000. A third still orders an article you discontinued two years ago, whose successor has a different pack size.

Units are where the silent errors come from. Ten pallets is not ten of anything you sell. It is 960 cases for one customer and 800 for another, because their pallet builds differ. Kilograms become sheets, metres become bars. A tool that hands a person the number 10 and the word Paletten has moved the work, not removed it. Ask where the conversion rule lives and who can read it without a ticket.

Ask what happens to a customer that exists three times in your ERP

Three entries named Müller GmbH, two of them dormant, and the order matches none exactly. Matching has to weigh several weak signals together instead of trusting one string, which is why the delivery address counts as much as the company name. If the answer begins with "first you clean up your master data", you have learned the timeline. It will not be clean before go-live, and a process automation project that waits for that never goes live.

3. What happens when something is wrong or missing?

The real test is exceptions. Ask what the system does with a missing quantity, a price variance or an ambiguous customer. A good answer: nothing is booked or sent below a confidence threshold. The case escalates to a person with a pre-drafted clarification. GeneralMind scores every decision and routes low-confidence cases to an operator with the draft ready to send.

Ask for the escalation numbers, not only the accuracy number. What share of orders reaches a person in month one, what does that person see, and how many steps does the correction take? A case that arrives as the raw PDF plus a red banner is not an escalation. It is the original job with an extra login.

Does a correction stick?

If an operator fixes the same customer's article mapping every Tuesday for six months, nothing is learning. Ask what a correction becomes: a permanent rule for that customer, on a list you can read. Klöckner books 81% of order lines with zero human edits. Oatly runs roughly 2,500 orders a month with 80% on Autopilot. Deployments typically start near 85% straight-through on day one and reach 93–95% within weeks, 90%+ autopilot in about six weeks. That is the usual curve, not a guarantee.

4. How deep is the ERP integration, and who maintains it?

Ask whether it writes back bidirectionally, whether it needs a migration, and who owns the interface over time. GeneralMind connects to SAP (including heavily customized ECC, S/4-ready), Oracle, Dynamics, Infor, NetSuite and 100+ systems via lightweight API, and owns interface maintenance so it isn't an internal burden.

Both directions carry weight. Reading means prices, master data, available stock, a credit block. Writing means an order that lands in the ERP looking like one your team keyed: same document type, same partner functions, the customer reference in the field your controlling reports on. Not a queue of drafts waiting for approval, which is the manual work under a new label.

If you run ECC with fifteen years of Z-tables, say so in the first call and ask what has to change on your side. The answer you want is nothing: no migration, no new module, no parallel data model. GeneralMind reaches your systems over a mailbox and a lightweight API, adapts to the interfaces each system already exposes, and maintains them itself. Ask who fixes that interface when your team makes a field mandatory in the spring. That answer decides whether digital order processing still runs in year two.

5. Where does the data live, and is every action auditable?

For German suppliers, EU data residency and GDPR aren't optional. Ask for hosting location, certifications and whether your data trains third-party models. GeneralMind is hosted in Frankfurt (DR in Stockholm), holds ISO 27001:2022, ISO 27701 and SOC 2 Type II, keeps data isolated and untrained, and logs every step (email in, classification, analysis, ERP sync) for a complete audit trail.

Push past the badges: ask for the certificate rather than the logo. It names the entity and the systems inside its scope, and the scope decides whether it covers the product you are buying.

Then ask to see one order's trail end to end, on screen, in the demo. Your auditor will not ask what the PDF contained. They will ask who accepted a quantity change from 500 to 480, against which rule, at what time. A record that cannot separate a human override from an automated decision is a log, not an audit trail.

Bonus: how is it priced, and how fast is go-live?

Outcome-based pricing aligns incentives. You pay per processed transaction, not per seat. Ask when costs start (GeneralMind: only at go-live) and how long deployment takes (weeks, with 90%+ autopilot typically reached in about six weeks as the system learns).

That model has a second effect worth naming: a vendor paid per processed transaction absorbs the cost of its own failures, while per-seat pricing pays the same whether the work left your team or not. Before you compare quotes, measure your baseline with two people and twenty real orders, timed from arrival to booked line with the clarification emails counted. Without that number no proposal is comparable.

Answers that should end the conversation

• Your IT owns the ERP interface. It gets built once, then ages.
• One accuracy percentage, with no split between fields read correctly and orders booked without a person. Only the second number gives you capacity back.
• A shrug about where order data is processed, or whether it trains someone else's model.

Run the demo on your worst week, not your cleanest day

AI-driven solutions get shown on a tidy order from a fictional customer. Pull thirty real emails out of a genuinely bad week instead: two scans, an order change buried in a reply, an order for a discontinued article, one from a ship-to address that is not in your master data, one in the wrong language for that mailbox. Send them cold and watch the run live, not a report afterwards.

Then score two things. How many of the thirty reached the ERP without a person touching them, and how long each took from arrival to booked. Everything else in the pitch sits downstream of those numbers. B2B sales technology gets bought in a forty-minute demo and lived with for five years. That is a bad trade for industrial suppliers running thousands of order lines a week. Spend the afternoon.