How to Build an AI Operating System · Construction Contractors Hub

How to Build an AI Operating System

AI is getting really good. Every day I’m blown away by what you can do in Claude.

If you aren't getting good results from it, I believe you are most likely doing one of two things wrong:

Insufficient context
Poor direction

Most people drastically underestimate how much background information you need to do a task well.

Writing a scope of work? You need the drawings, project specifications, previous templates, package register and a list of previous variations from past projects.

Then you need to define exactly how to do it. What goes into each section, how to populate the pricing schedule, systematically check that all head-contract requirements are transferred to the sub-contractor.

Generic AI is generic. It doesn't know construction process, and it knows nothing about your business — your rates, your subbies, your contract positions, how you run a job.

An AI Operating System solves this problem.

This week, I'll cover what it is, the parts and how to build it.

Check out ContractorOS. We have built most of this for you. Pre-built skills, the business brain structure and weekly calls to help you implement it.

Purpose-built AI workflows you download and install. We run weekly workshops and offer unlimited one-on-one tech support to ensure you get the most out of AI.

What an AI operating system actually is

A system that lets you get genuinely useful work out of AI — not a chatbot you ask questions.

Out of the box, AI has three limits: it's generic, it doesn't know the construction process, and it has none of your business background.

The fix is two things: give it the right process (a workflow) and the right data (context).

Fundamentally, an AI OS is data + workflows. Everything else is plumbing to serve those two.

One mindset shift before the parts: use AI as a data transformation tool, not a coworker. It doesn't think like a senior estimator. It moves data you already have into the shape you need.

The two ways you use AI

Ad hoc — one-offs. "What are the concrete specs on this drawing?" Needs good background context sitting where AI can reach it.
Workflow — repeatable tasks you run on every project. Conceptual estimate, go/no-go, payment claim. Needs that context plus a defined process — the exact steps, every time.

The OS serves both. The data layer powers ad hoc; data + workflows power the repeatable stuff.

The eight parts

Your existing software stack — accounting, project management, estimating, takeoff, programme. The OS doesn't replace these. It sits across them.
The AI tool — Claude, ChatGPT or Gemini. The engine.
Your project folder structure — SharePoint or Google Drive, with AI connected to it.
Connectors (MCP) — the cables between the engine, your software and your data.
Workflows, stored as skills — your repeatable processes, captured once. Keep copies in GitHub; bundle them into a plugin so the whole team runs the same process.
Triggered and scheduled tasks — workflows that fire on a clock or an event, not just when you ask.
Business data — cost library, contract positions, subbie directory, lessons learned. Somewhere central and accessible (ours is in Notion).
The config file (claude.md) — how the AI behaves and where everything lives.

The Business Brain

The rule: business data lives in one central place; project data lives in the project folder.

The brain holds what spans every project and changes slowly:

Business background — who you are, team, markets, software stack
Cost history — what your jobs actually cost, by activity
Contract positions — your standard acceptable position on each clause
SOPs and templates
Subcontractor and supplier directory
Lessons learned
Projects index — one summary per job

Write it for two readers: a new hire on day one, and the AI. If a new starter could run the business from it, so can Claude.

This is what makes the difference between an estimate built on invented rates and one built on your actual cost history.

The build order

Stand up your business data somewhere accessible.
Set up your project folder structure and connect AI to it.
Write the claude.md — navigation, behaviour, guardrails.
Build your first workflow as a skill. Run it manually 3–4 times first — only systematise what already works by hand.
Add connectors as workflows need them.
Put skills in GitHub, bundle into a plugin for the team.
Add scheduled routines last.

Start with one workflow and one project. Don't build the whole thing on day one.

The Bottom Line

An AI OS is data + workflows — everything else is plumbing.

Give AI your process and your data, and it stops inventing rates and starts producing work you can defend. Build around its limits: it transforms data, it doesn't replace your judgment. Do the deep work yourself.

One workflow, one project, then expand.

4 comments