CCSI is an Auckland-based commercial structured-cabling contractor: data, voice, fibre, wireless and low-voltage systems for commercial fit-outs, schools, and infrastructure projects. This is the evidence behind the Automation Integrator approach, tested against real customer money and real deadlines for months, not a demo week.
Commercial cabling tenders in Auckland are won and lost on price and speed of response. Plans arrive on short notice, and the contractor who turns a competitive quote around first, without a pricing mistake, tends to be the one who gets the next call too.
Like most trade and services businesses, CCSI's owner was running sales, estimating, project delivery, finance, and marketing at once, usually before 7am and after the crew had gone home. The mechanical load behind each of those jobs, reading and re-reading the inbox, retyping numbers from a PDF into a pricing sheet, chasing the same overdue invoice, was consuming the hours that should have gone to winning and running work.
The brief was specific: use AI to take over that mechanical load, without taking judgment, pricing decisions, or client relationships out of the owner's hands.
Communications went first, since it was the single biggest and fastest-proven win. Each function only started once the one before it had proven itself in live use.
Replies to every new email (never sends), matched to sender context, pulling the right price, the right file, the right tone. Replaced a paid third-party drafting tool once it proved it could pull a customer-safe sell price while refusing to leak the cost price in the same file.
An industry tender portal checked twice weekly and relevance-scored; a website "instant estimate" tool lets prospects self-serve a rough price, capturing leads with zero manual involvement. A daily scan raises won work into the job register and pipeline automatically, and a weekly scan of main-contractor award notices flags and files the ones relevant to the business.
Reads a legend, learns a drawing set's symbols, counts every match across every sheet, produces a marked-up proof PDF plus a filled pricing sheet. Benchmarked against 11 historical jobs before ever touching a live tender.
Debtor chasing, cashflow forecasting, payment run prep, supplier statement reconciliation, progress-billing invoice drafting, purchase order raising, contractor invoice allocation, supplier invoice processing, and a monthly audit that every completed variation actually got invoiced, all advisory or draft-only, none moving money autonomously.
Every outbound proposal is scanned from sent mail and logged into the CRM automatically, the one automation allowed to write without a human approving each write, because a wrong CRM entry is cheap to fix. The weekly award-notice scan moves a customer's stage forward automatically when they're reported as awarded work.
A purpose-built web app: office side for document management, crew side needing no login, just a link or QR code. Crew-facing views are stripped of all pricing information in code, not just policy.
Brand positioning locked once, then a weekly automation builds and schedules the next week's social content across every channel. Approve-first by design, no autonomous posting or spend.
A dashboard shows every automation's live status in one place, framed as a team roster rather than a technical report. A daily health-check emails a flagged alert if anything's broken. A message bus now lets a won tender wake the job-raising and PO automations directly, and a nightly review reads back through the day's work and proposes fixes or new standing instructions for approval.
The first attempt at scheduling these automations used the AI platform's own built-in cloud and desktop schedulers. It mostly worked, but desktop-scheduled tasks only ran while the app stayed open, and the cloud scheduler occasionally missed a fire with no obvious cause.
Everything was migrated onto the operating system's own native task scheduler, decoupling the automations from any app needing to be open. A separate lesson landed the same week: the AI platform's own cloud scheduler had a real, unexplained failure mode, every scheduled automation on it silently stopped firing account-wide for a full day. A strong argument for not depending entirely on any single cloud scheduler for anything time-sensitive.
A second always-on machine was trialled as redundancy and reversed: the managed cloud PC deallocated on every disconnect, silently killing every scheduled task on it the moment nobody was actively connected. The fleet runs from the original machine today, no fallback, a known and still-open single point of failure rather than a solved one.
The proposal that started this build is a standalone read, no CCSI-specific detail required.
Tell Us Where Your Time Goes