The three workflows a construction company should hand to AI first
Accounts payable: receive, read, route, chase
Five hundred to a thousand supplier invoices a month, previously judged one by one by a quantity surveyor for site, project type, purchase order and approver. Now a dedicated mailbox is monitored, OCR plus a large model extract the fields, files go to OneDrive by site, POs are matched, rule-compliant invoices are submitted to ApprovalMax and the rest are queued for a person by cause. In one accounting month 730 of 798 documents were routed automatically, 91.48%; approval and payment still happen in ApprovalMax and Xero, by people.
Site hours: from one sentence to a cost dashboard
Forty-odd workers across fifteen to twenty projects, with hours previously recalled by foremen after work and typed into Excel. Now each site has a Lark group; the foreman posts a sentence or a photo and @-mentions the agent, which writes the daily record and hours detail; a Base automation applies pay rates into the monthly payroll summary and the owner reads labour cost by project. Kickoff 8 April, company-wide 15 June, and at the 27 June handover 20 projects, 334 records and 10,000+ hours were in the system. Payroll never passes through a language model.
Enquiries and quotes: chase the details, then price
A roofing company’s enquiries used to be jotted anywhere, quotes went out missing measurements and photos, and days were lost to back-and-forth. Now the agent chases a checklist of measurements, photos, address and timing, generates the project card and follow-up scripts, archives site photos and reminds on warranty expiry. The estimator still sets the price.
The shared boundary: AI prepares, people decide
All three projects follow one rule: for approval, payment, pay-rate changes and quote commitments, the AI prepares and checks and a person presses the last button. Data stays in the client’s own environment (a Mac mini in the office or the client’s own Lark workspace), permissions are isolated by role, and who sees amounts versus only their own queue is configurable.
Two of these workflows have full written case studies: three months to 91% invoice automation and site timesheets from Excel to a chat group. For what projects like these cost, read how much an AI digital employee costs in New Zealand; for all eight anonymised projects see AI implementation case studies. BEE Sigma connects rather than replaces, implements rather than theorises, and stays involved after launch.