Case study: reading 7,369 order emails a month for a UK glass manufacturer

The problem

The business makes glass units for window and door companies. Its trade customers send orders by email, and no two customers send them the same way. When the project started, there were at least six distinct formats, including:

• order sheets generated by the customer’s own software
• formal purchase orders with product codes and delivery addresses
• typed tables with hand-drawn sketches and handwritten notes added
• handwritten faxes

Spreadsheets, Word documents and photographed order forms followed.

Each one uses its own words for the same thing. “Soft Coat”, “S/Coat” and “Planitherm” can all mean the same glass. Staff had to read every email, tag it in Outlook, work out what was urgent, count the units, find the delivery date, and then key the order into the production system.

The surprise, once we talked it through with the staff, was that typing wasn’t the main pain. It was knowing what to do first. With hundreds of emails a day, the hard part was spotting which ones were orders, how big they were, when they were needed, and which were urgent.

What the system does

Every email sent to the orders address arrives in the system automatically, with its attachments. Then, without anyone touching it:

1. It works out who sent it. The sender’s email domain is matched to a customer. If it can’t be matched, it waits for a person to assign it.
2. It decides what it is. An AI model reads the email and attachments and sorts it: order, enquiry or something else.
3. It reads the order. How many units, when they’re needed, and whether it’s urgent. It reads typed PDFs, spreadsheets, Word files, photos and handwritten faxes.
4. It works out the dates. Lead times are counted in working days. A requested date that falls on a weekend, or is too soon, gets a Date Check flag so someone confirms it.
5. It reads every line. Each glass unit is pulled out with its sizes and specification, matched to the product catalogue, and flagged if it’s oversized.
6. It queues the order. The order goes straight to the first-entry queue with everything a member of staff needs, already sorted by priority.

How it’s checked

Nothing goes to production on the system’s word alone. Every order passes through two checks by two different people, and the system enforces that the second check can’t be done by the person who did the first. Every change is recorded in an audit trail.

When the system can’t read something fully, the order still goes through, carrying a warning that says why. A system that never asks for help is one that’s quietly guessing.

It learns from corrections

Two things make it better every week:

• Product matches. When staff change which product a line was matched to, that pairing is remembered. The next time the same description turns up, it’s matched automatically.
• Dates. When staff correct a date the system read, the correction is kept against that customer. The most recent corrections are shown to the AI as worked examples the next time that customer’s order is read. Staff can also write plain-English notes for a customer, such as which box on their form holds the real delivery date.

Keeping the running costs down

AI reading costs money per read, so the system uses the right model for each job. Typed PDFs go to a faster, cheaper model. Photos and scans, where handwriting is the problem, go to a stronger one. Photos are resized before they’re sent, because the model would shrink them anyway.

In the 30 days to 7 September, every read the system made, from sorting emails to pulling out each order line, cost about $425 in AI usage. The dashboard shows the running cost today, this week and this month, so there are no surprises.

How it was built

Not in one go. Work started in February 2026, and since then over 200 improvements have been made, each one written up in plain English for the staff who use it. Plenty are the kind of thing only the people using it would spot: a customer whose dates kept being misread, a form whose empty tick boxes were being read as ticked, sizes like 667.5mm that needed rounding down because glass is always cut to the size below.

It runs with over 800 automated tests, so a fix for one customer’s paperwork doesn’t break another’s.

What’s next

Staff still key finished orders into the production system. The next step, now in testing, is an import file that sends a checked order straight into it. When that’s live, the keying goes too.

What it means for the team

Before, every email had to be read before anyone knew what it was. Now staff start from a prioritised queue: what’s urgent, what’s big, what needs a date confirmed. Their time goes on checking and on the orders that need a person, not on reading every email to find out what it is.

If your business has a version of this problem, this is how it works in general, and this is what it costs.