GlassCo Metro makes and fits windows, glass and shower screens across Western Australia. Every day, builders and suppliers sent them a pile of emails and delivery notes that someone had to read and type into the computer by hand. We built two AI helpers that read all of it, do the typing, and pass anything tricky to a person. They work on their own, around the clock, without anyone having to start them.
Both of these turn up as ordinary emails and PDF files — written for people to read, not for computers. That is what makes them hard. A normal computer program needs information to be in exactly the same place every time, so it breaks the moment a supplier moves a column or a builder changes their wording. A person can just read it and understand. That is the gap we needed to close.
When a house is ready for its windows, the builder emails GlassCo to say “come and do this part now”. Someone had to read each email, find the right job on the computer, check the job was actually up to that stage, and then set up a to-do list for the right people. Every builder writes their email a little differently.
Every delivery arrives with a note listing what is in the load. Someone had to open each one, read every line, compare it against what was ordered, spot anything missing or short, and tick it all off on the computer. Every supplier lays their note out differently, and some run over several pages.
None of it is difficult work. It is just endless, and it has to be right — a mistyped line means a crew turns up to a house that is not ready, or the stock on the computer no longer matches what is really in the warehouse.
Neither of these is a chatbot you talk to. Nobody types anything into them. Each one wakes up on its own, picks up whatever has arrived, and works out whether it understands it well enough to act. If it does, it finishes the job. If it does not, it stops and tells a person why.
Wakes up every hour and checks its own mailbox for new emails from builders.
It handles six different kinds of request, from window frames to shower screens. Each one has its own rules, and it knows which rules apply.
Wakes up every hour and checks its own mailbox for delivery notes from suppliers.
It copes with several different suppliers, each with their own layout, including notes that run over several pages.
Both helpers follow the same seven steps. The important idea is that the AI is only used for the one thing it is genuinely good at — reading writing that does not follow a fixed pattern. Every decision that has to be correct every single time is handled by plain, fixed rules instead. The AI reads. The rules decide.
Nobody presses start. A clock inside the system switches the helper on every hour. It signs in to its own mailbox, exactly like a staff member would, and picks up whatever has arrived.
Delivery notes arrive as PDFs, which are really just pictures of a page. Special software reads the words and the table lines off that picture, so the helper knows what the columns and rows say. Builder emails are read straight from the email itself.
The AI decides what kind of document this is — and, importantly, gives a score for how sure it is. A wrong guess here would send everything downstream wrong, so this step exists to catch it early.
The AI pulls out the specific facts the job needs: the job number, the date, the person, the products, the quantities. It also scores how sure it is about each one, so a shaky reading can be spotted rather than quietly trusted.
This is the part that is not AI at all. Ordinary, fixed rules take over and check everything against the real records: does this job exist, is it at the right stage, do the delivered amounts match the order? Rules like these always behave the same way, so the answer is never a matter of opinion.
Nothing gets written down until every check has passed. Then the helper sets up the tasks, attaches the plans, marks the delivery as received and updates the stock — the typing a person used to do.
Every document produces an email: done, could not do it, or needs a person. At the end of each round it sends a summary listing everything it touched, so staff can see its work without digging through anything technical.
Software that changes real business records while nobody is watching has to be trustworthy in a way a demo never does. Getting the AI to read a document is the easy bit. The real work is everything built around it to keep it inside its lane. These are the safeguards that make it sensible to leave running.
Every time the AI reads something, it also gives a score for how confident it is. If that score is too low, the work stops there and a person is asked to look. The AI does not get to decide whether it is sure enough — that line is drawn in the code, and the AI cannot move it.
When something does not add up — a short delivery, a name that does not match, a document it cannot make sense of — the helper sends an email saying exactly what is wrong and what a person needs to do. Stopping and asking counts as doing its job properly.
One kind of request is never handled automatically, even though it could be. Fitting shower screens has to be booked around other trades on site, and that takes human judgement about things the computer cannot see. So the helper flags those and steps back. We chose what should be automatic, not just what could be.
The AI never decides whether a job can go ahead, what tasks to create, or whether a delivery is correct. Those decisions are made by ordinary fixed rules that behave the same way every time and can be read and checked by a person. The AI only turns messy writing into tidy facts.
If another system is busy or offline, the helper waits a moment and tries again, a few times, before giving up. If it does give up, it stops that one document and tells someone. What it will never do is write half an answer and leave the records in a muddle.
There is a second set of code whose only job is to check the first. Every time we change anything, it re-runs the tricky cases — short deliveries, odd product names, missing details — and tells us immediately if something that used to work has stopped working.
At the end of every round, the helper sends a plain summary: here is each thing I looked at, here is what I did with it, here is what I passed to you and why. Nobody needs to be technical to check up on it.
The helper is given the keys to one mailbox and nothing else. It cannot read anyone’s personal email or wander into other parts of the business. Each connection it has is limited to just the part it actually needs to do its job.
GlassCo Metro makes windows, but nothing we built is specific to windows. Underneath, the skill is taking a steady stream of paperwork that nobody has time to read and turning it into correct entries in the system a business depends on — while being clear about what software is allowed to decide on its own. That same skill applies anywhere the admin is real and the paperwork is messy.
We build these the way any serious software gets built: written carefully, tested automatically, released properly, and explained well enough that the people relying on it understand its rules. That is the difference between trying AI out and actually being able to depend on it.
If people are spending their day reading emails or forms and typing the same details into another system, there is usually a helper worth building. Let's talk about what should be automatic — and what should stay with your people.
Get in touch →