Automating orders that arrive by email and Excel

The AI reads every incoming email and its Excel or CSV attachments, recognizes which ones contain an order, and extracts the line items: product, code, quantity, customer and delivery references. It normalizes the data into one consistent format, flags urgent requests, and preps the entry in your ERP. Nothing is recorded until a person confirms it.

In practice it replaces the repetitive, error-prone part: open, interpret, retype. It doesn't decide on its own what to ship or at what price. It turns messy messages and spreadsheets into clean, ready-to-post order lines, so your sales desk reviews and confirms instead of keying data in by hand.

The flow has five steps. One: the AI monitors the inbox and catches messages containing orders. Two: it reads the message body and any Excel/CSV attachments. Three: it extracts and normalizes the line items (product, quantity, references). Four: it flags urgent, missing, or inconsistent data. Five: it preps the ERP entry, which a person confirms before anything is recorded.

The system handles the fact that every customer writes differently: some include codes, some only a description, some attach a file, some write in prose. The AI maps it all to the same schema. When a value is ambiguous or absent it doesn't guess: it flags the item and leaves the call to your operator.

It integrates with your email inbox (where orders arrive), with the Excel and CSV files used as attachments, and with your ERP or management system where line items get recorded. Depending on the architecture it also connects to your CRM, price lists, and customer records to match codes and references. The goal is to fit your existing flow without replacing your ERP.

The ERP connection can run through APIs where they exist, or through structured file import when they don't. Every project starts with a diagnosis of your real tools: which inbox, which attachment formats recur, how your ERP is built. From there we define the most robust, least invasive integration.

The final sign-off. The AI preps the order lines, but they're recorded in the ERP only after an operator reviews and approves them. For higher-risk cases — large quantities, pricing, new customers, inconsistent data — the decision explicitly stays with a person. The AI speeds up and organizes the work; control over what enters your systems stays human.

This guards against two opposite errors: the lost order and the wrong order posted automatically. The operator sees a screen that's already filled in and clean, with urgent items highlighted and doubts flagged, and confirms with a click when everything checks out. Time shifts from typing to checking, where human judgment actually matters.

The most direct KPIs are three: orders not lost (nothing stays buried in the inbox), turnaround time (from receipt to a recorded order), and transcription errors (wrong codes and quantities). You can also track operator time per order and the share of urgent requests handled same-day. The actual figures depend on your volumes and should be measured before and after.

The idea is to start from a snapshot of your current state — how many orders a day, how long they take to enter, how many downstream errors — and compare it after go-live. We don't promise a percentage: we promise to make the process measurable and to work on your numbers, not on industry averages.

It's not worth it if you get very few orders a day and they're already well structured: the time saved won't justify the project. Nor if orders are almost all placed by phone or in person, or if each request is a complex, negotiated quote rather than a repeatable order. In those cases the bottleneck is elsewhere and automation adds little.

Automation pays off when there's volume and repetition: many orders, many formats, a lot of manual retyping. If you're not sure you're in that range, the initial diagnosis exists precisely to tell you honestly — including telling you not to do it. A reasoned no beats a project that never pays back.

Your order and customer data stays yours and is not used to train models (zero-training). EU data residency is available on request. Where the architecture allows, encryption in transit and at rest and least-privilege access with logging apply. We don't make absolute promises: the security level is defined around your specific case.

In practice, emails and attachments are processed to extract order lines, not to feed third-party systems. Decisions about where data lives, who can access it, and for how long are agreed during the design phase, in line with the GDPR and your own internal policies.