Hundreds of orders a day — created by AI
- Jul 8
- 4 min read
Customers email their orders as free text — no article numbers, approximate product names, special requests. At a Swedish ready-meal producer, customer orders are now created automatically in CloudOffice: the AI reads the emails, interprets the content and only leaves the uncertain cases to a human. Here is how it was done.
The starting point: repetitive, time-consuming and error-prone
The company delivers ready meals and catering to shops and cafés across Sweden. Orders arrive by email, and every email required manual entry into the ERP: find the customer, interpret the product names, key in lines, dates and requests. Time-consuming in everyday work — and a real risk of errors in high season, when volumes peak.
Together with an external AI specialist, two areas with the greatest potential were identified: order email processing and forecasting of order volumes for the in-store assortment. They started with the order emails — everything needed was already available in CloudOffice's open API.
The solution: AI reads the emails, orders wait as drafts
The flow looks like this:
Emails are collected from the company's order inbox. The AI reads every email and extracts the order information: customer, articles, quantities, delivery day and other requests — delivery times, allergens and anything else the customer wrote in free text.
The data is matched against the customer and article registers from CloudOffice — even when the customer writes approximate product names without article numbers.
A customer order is created via the API — always as a draft. Nothing is posted live without the company's own approval. Customer, delivery date, route and vehicle are filled in automatically (the route is taken from the customer card, or from the latest delivery if none is set), and the customer's original email text is saved in the order's comment field.
If the AI is confident about the order, it is left clean — staff can check the lines directly in the order list and approve the order without even opening it. If anything is uncertain, the order is marked [REVIEW] in the reference field, and the comment states exactly what the AI is unsure about.
Built-in safety checks
From day one, the automation had more controls than the manual process ever had:
Duplicate check. When an order is created, the system looks for similar orders — same customer, same reference, same articles — and flags the risk of double entry.
Date check. Customers sometimes make mistakes in their emails: the wrong year (2027 instead of 2026), or "Saturday June 10" when the 10th isn't a Saturday. Such orders are flagged for review, as are delivery dates unreasonably far ahead.
Traceability in the inbox. Every email the AI has processed gets an "AI-processed" label. Read + label means the order is already waiting in CloudOffice. Unread + label means the AI could not create an order — the email needs a human. Staff keep working in their usual email environment, with full visibility of what is done and what is pending.
The result
Hundreds of orders a day are now handled fully automatically. Most orders land pre-filled in CloudOffice, staff review instead of typing — and the built-in checks catch errors that previously could slip through. Many hours are saved every day, and the customer's order manager now shows off the day's automatically registered orders with some pride.
The next level: forecasts instead of gut feeling
Part two of the project concerns order volumes. The company has shelf agreements with stores: the shelf must be kept stocked, but the producer carries the risk of waste. Too much means returns; too little means empty shelves and lost sales. Adjusting orders manually — one order line at a time, per store — was something staff simply couldn't keep up with.
The first step is an analysis report built from the sales history in CloudOffice: per store and article it shows historical delivery days, delivered and returned volumes, and return rates with color coding — so buyers quickly see where to act and can adjust volumes in bulk instead of line by line.
The next step is already largely built: a forecasting engine that suggests the order volume per store, article and delivery day. The model learns from sales history, the calendar (holidays, long weekends) and the weather — weather sensitivity varies dramatically between stores. It also accounts for profitability: with high gross margins, a lost sale costs more than a return, and the model is calibrated per article accordingly. A single model covers all stores and learns from similar stores — so even new stores with little history get sensible suggestions. In tests, the forecasting engine already performs on par with an experienced buyer. Rollout is being done step by step, as the data history grows.
How it's built — and why it went fast
The whole solution is a standalone program that talks to CloudOffice through our open API — the same API documented at developers.cloudoffice.se. Emails are collected via the company's regular email environment, weather data via an open weather service, and every ten minutes the inbox is checked for new orders. No special adaptations of CloudOffice were needed: customer and article registers, customer orders, sales history and agreements — all of it was already accessible through the API.
This is exactly how we see AI in the ERP: CloudOffice is the reliable base where all data comes together — and the openness lets AI solutions plug in and do the work.
Want to do something similar?
Do you have a workflow with lots of manual steps — order emails, purchasing, forecasts? Book an AI demo and we'll show this case live and discuss what could be automated at your company.
Company details and references available on request.
Comments