Build Your Own AI Assistant for a Contracting Business (Step by Step)

Reviewed by the Contractor AI Shop research team · Last updated Nov 5, 2026 · How we research

Short answer: You can build a working AI assistant for your office without hiring anyone. It takes four parts: an AI model (one you run on your own computer, or a cloud AI's free or paid plan), a small folder of your own company documents, a free assistant app such as OpenClaw or Hermes Agent that you can message from a chat app, and one automation built in n8n that runs on a schedule. Start with one job, test it on questions you already know the answers to, and only add more once it gets those right. We haven't timed this setup end to end, so we won't guess how long it takes.

At a glance: what you need

What this assistant won't do: know anything you didn't give it, or be right just because it sounds sure. It can misread a document or answer a question its files don't cover. Keep a person in charge of anything about money, scope, contracts or safety.

1. Pick one job for it

Write one sentence before you install anything. For example: "Answer office questions from our standard documents and name the file the answer came from." Good first jobs are the questions your office answers over and over: what's in our standard exclusions, where's the safety program section on fall protection, what does our subcontract say about retention. Keep it to one job for the first two weeks.

2. Choose where the model runs

• Private (documents stay on your computer): install Ollama and pull a model. Our guide to chatting with your specs locally covers the install, and which local models fit which hardware helps you pick one. On a Mac, see local AI on a Mac Studio; for an office server, see the DGX Spark setup guide.
• Cloud (easier start): use a cloud AI's free or paid plan. Free plans have limits that change, and before you send company documents to any cloud service, check that provider's data settings for your plan.

3. Install the assistant app

Both of these run on your own computer, connect to the model you picked, and let you message the assistant from a chat app instead of a browser tab. Use the vendor's own installer and docs; we link rather than copy their commands, because install steps change.

• OpenClaw: free and open source (MIT), for Mac, Windows and Linux. The vendor says there's no paid tier and no hosted service. Start with OpenClaw's getting-started docs; its Ollama provider docs cover using a local model. Chat channels in its docs include Discord, Google Chat, iMessage and Matrix (its docs also cover Slack, Telegram, WhatsApp, Teams and others).
• Hermes Agent (Nous Research): free and open source (MIT), with memory and skills. The Hermes quickstart offers a desktop installer for Mac and Windows or a command-line install. Run hermes setup to get started, or hermes model to pick a provider, including a local or self-hosted endpoint. Once a normal chat works, hermes gateway setup connects it to Telegram, Discord, Slack or another platform. The quickstart's own advice is worth following: get one clean conversation working before you add anything else.

Not sure which to pick? Try whichever one supports the chat app your office already uses. Both are listed in our agent tools category with other options.

4. Give it a clean folder and its own logins

This is the step people skip, and it's the one that matters most.

• Make a dedicated folder with only the documents for its one job: standard exclusions and qualifications, safety program, a sample subcontract, your company FAQ. Leave out client pricing, bank details, HR files and anything you wouldn't email to a stranger.
• Never point an assistant at a general folder like Downloads or your whole desktop. Security researchers recently showed an AI coding agent being hijacked by a malicious instruction file sitting in a Downloads folder (our brief).
• Give the assistant its own accounts, not a person's email or logins, so you can cut it off without locking anyone out.
• Leave approval prompts on. If the app asks before it runs a command, edits a file or sends a message, keep it that way.

5. Give it standing instructions and test it like a new hire

Paste this as the assistant's instructions (system prompt):

You are the office assistant for a [trade] contractor. Answer only from the documents provided. For every answer, name the file and section you used. If the answer isn't in the documents, say "Not in our documents" and suggest who in the office might know. Never make up prices, dates, contract terms or code requirements. Keep answers short and in plain language. Before answering a question about scope, money or safety, ask one clarifying question if anything is unclear.

Then test it: ask three questions you already know the answers to, plus one whose answer isn't in the folder. It should name the right file for the first three and say "Not in our documents" for the fourth. If it answers the fourth anyway, don't use it for real questions yet. Tighten the instructions, trim the folder, or try a larger model.

6. Add one automation with n8n

An automation is what makes this an assistant rather than a chat window: it does a job on a schedule without you asking.

1. Self-host n8n's free Community Edition. The n8n hosting docs cover the options; the Docker install is the most common. If nobody in the office wants to run a server, n8n Cloud Starter is €20/mo billed annually (Pro €50/mo), per n8n's pricing page on Nov 5, 2026.
2. Build one workflow: a Schedule Trigger each morning, then an AI Agent step that reads one thing you choose (for example, a shared folder of yesterday's daily reports), then a short summary sent to your chat or email.
3. Point the AI step at your model. For a local model, n8n has an Ollama Chat Model node; for a cloud model, add that provider's key.
4. Have it send to you first, not the whole office, until you've read a week of summaries and trust them.

Comfortable with Python? CrewAI lets you script a small "crew" of AI agents that each take one role, for example one reads a spec section and another drafts the RFI list. It's MIT licensed and can use a local model through Ollama, per its README. Its README also says anonymous telemetry is on by default; it can be turned off by setting OTEL_SDK_DISABLED=true.

7. Review after two weeks

Keep what it answered right. When it gets something wrong, fix the document behind the wrong answer before you blame the model. Add a second job only once the first one is boringly reliable.

When it's worth paying (or getting help)

A self-hosted assistant only works while its computer is on, updated and backed up, and someone in your office owns that. If nobody does, a hosted option can save the upkeep: n8n Cloud, a paid cloud model, or a hosted plan from the assistant's vendor (model providers and hosted services are priced separately from the free app). OpenClaw has no hosted tier. Start self-hosted, and pay once you know which job the assistant actually does for you. Weighing outside help instead? Our AI consultant vs doing it yourself checklist covers when hiring someone makes sense.

FAQ

Related guides

• Chat with your specs and bid docs locally
• Which local AI models fit on which hardware
• A free AI estimating setup you can build yourself
• Free AI tools for contractors (2026)
• AI consultant vs doing it yourself