Where they started
A Sydney construction business running Procore for project management, with years of work behind it. Every job had left a trail: tender submissions, subcontractor quotes, site notes, variations, delay notices, supplier correspondence. About 10,000 documents, almost all of them PDFs, spread across job folders and inboxes.
Procore held the live projects well. It did not hold the answers to the questions that kept coming up. Why did we go with that supplier over the other one on the last hospital job? What caused the six week delay on the school project? What did we price that scope at two years ago? The answers existed, but finding them meant someone who was there at the time, or hours of digging, and often both.
What we built first
The base layer is what we call a company brain. We took every document, current and historical, and converted it into a structured, machine readable form. We kept the full content rather than summaries, with the job, date, document type and parties attached.
All of it runs on a private server the business rents. The model that reads and answers questions runs on that server too. No document, and no question about one, goes to a public AI service. For a builder whose files contain subcontractor pricing and client commercial terms, that was the condition for doing it at all.
Every answer cites the documents it came from. A project manager can check the source rather than trust the summary. Procore stayed exactly where it was as the system of record, and nothing was migrated out of it.
What it made possible
Years of decisions became askable. Why a particular supplier was chosen over another on a past job, with the quotes and the email thread that decided it. That is the procurement question every project manager asks and nobody was able to answer quickly. What delayed a specific project, with the site notes and notices in order. How a similar scope was priced last time, and what it cost.
Ten staff use it. A question that took hours of digging now takes seconds. The people who had carried that knowledge in their heads stopped being the bottleneck for it. New project managers can find out how the business handled something before, instead of guessing or interrupting a director.
Then the automations
With the documents structured and the cost history reachable, workflow automation became straightforward, because the data the automations needed already existed in a usable form.
The one the team uses most is estimating over WhatsApp. An estimator or site lead sends a job description, and photos or a scope where there is one, to a WhatsApp number. The system drafts an estimate from the business’s own cost catalogues and past pricing, in their template, and sends it back for review. A first draft that took hours at a desk now takes minutes, and a person still owns the number.
Further automations followed on the same foundation. Each one was cheaper than the last because the business’s own information was already machine readable.
Why the order mattered
Most construction businesses ask for the automations first. We did the document layer first because every automation worth building for this client depended on the same thing: reliable access to their own history. Estimates from their cost catalogue, supplier decisions from their tender files, delay analysis from their site notes. Build the layer once and each workflow on top is a small job. Skip it and every automation has to solve the data problem again.