Case study: a 24/7 AI assistant for a busy NZ construction company

The brief

A busy New Zealand construction company came to us with a familiar problem. Their day-to-day was running through one person — the project manager — who held most of the client history, job status, supplier details, and quoting context in his head and across a sprawl of spreadsheets and old emails.

When he was on site, in a meeting, or simply offline, the rest of the team got stuck. The owner could not quickly pull up the status of an active job. Office staff could not check whether a client had outstanding paperwork. Everything had to wait.

They did not want to hire another senior person. They wanted a system.

What we built

We deployed a private, internal AI assistant for the company. The core stack:

• Open WebUI as the AI front-end, self-hosted on the company's own infrastructure
• Read-only connection to the internal database, so the assistant could answer questions about clients, projects, and history accurately — without ever being able to change anything
• Telegram bot for the owner, where he could ask quick questions from his phone on site
• Discord for the office team, with role-based channels — staff only see information they are already authorised to see

The assistant runs continuously. It does not get tired, it does not go on leave, and it does not forget what was said last week.

How it actually works

The owner sends a message in Telegram: "What's the status of the Albany job and who did we last invoice?"

The bot pulls the latest information from the read-only database, formats a clear answer, and sends it back in seconds. No spreadsheets, no asking around, no waiting for the project manager to be free.

Office staff have their own Discord channel, scoped to the data they have permission to see. Quotes assistants ask about historical pricing. Admin staff ask about outstanding documents. The same underlying assistant, different visibility — controlled by Discord roles tied to database-level access rules.

Security and operations

A self-hosted AI assistant with database access is only useful if it is genuinely safe. We treated security as a feature, not an afterthought:

• The database connection is strictly read-only — the assistant cannot insert, update, or delete anything
• No company data leaves the network — Open WebUI is hosted in their environment, with carefully chosen model endpoints
• Access is role-based — the Telegram bot is locked to the owner's account; Discord channels enforce visibility per team
• Every conversation is logged and monitored, so abnormal access patterns surface fast
• The bot is health-checked and runs 24/7, with alerts to us if anything goes down

The result

The team has stopped relying on a single person to remember everything. The owner gets answers from his phone in seconds. Office staff stop interrupting the project manager for routine questions. Domain knowledge that used to live in one head now lives in a process.

What changed is not just the speed of answers. It is that the business no longer depends on one person being available.

This was the company's first real AI integration — and it solved a problem they had been working around for years.

What this means for other SMBs

If your business has someone who is the "single point of knowledge" — the long-tenured operator who remembers every client, every quirk, every history — you have the same risk this construction firm had. A pragmatic, secure AI assistant connected to your existing data, accessible where your team already chats, is now a realistic project. It does not need a year. It does not need a CTO. It needs the right scope and a careful build.

Want to talk about whether this would fit your business? Book a free 30-minute AI assessment — we will tell you honestly if it is a fit, and what it would cost.