Start at the easiest level
Most owners should start at Level 1 and only move up if they hit a wall.
Level 1: upload files to a custom assistant. Drop your documents into a Claude Project, Custom GPT, or Gemini Gem and ask questions against them. That is it. You are already doing the thing people call RAG; on Claude it switches on by itself and scales as you add more. No build, no engineer.
Level 2: a purpose-built knowledge-base tool. When you want a chat widget on your website or in Slack that answers from your docs, a dedicated tool adds that front end. More setup, more reach.
Level 3: a custom pipeline on an automation platform. Tools like n8n can pull documents from a shared drive, read scanned files, and keep the knowledge synced. This only pays off when you need live updating or high query volume. For most 10-to-75-person shops it is overkill.
The privacy checklist (read before you upload anything)
This is where a service business can get itself in trouble.
- Know the default. On the free and personal tiers, the big tools may use what you upload to train their models unless you turn it off. Business and team tiers are off by default.
- Use a business tier for client data. Anything with customer names, records, or personal information goes in a business or team account, never a free personal one.
- Retention is not the same as training. Even with training off, data may be held for a window for abuse monitoring. True zero-retention is usually an enterprise-only setting.
- Check your client NDAs first. Many contracts now require you to disclose the AI tools you use and forbid feeding their material to one that trains on it. Uploading client documents can breach an agreement you already signed. When in doubt, get legal eyes on it.
- Some things never go in a shared tool: SSNs, health information, and the confidential terms of contracts.
Where it breaks, and how to keep it honest
- It does not auto-update. Editing the original file does not refresh the assistant. You have to re-upload. Put a monthly re-upload on the calendar.
- Keep one clean copy. Delete old versions so the AI cannot quote last year's price sheet. Put a last updated date inside each document.
- Messy files give messy answers. Scanned PDFs and complex tables confuse it. Use clean text and simple headings.
- It can still be wrong. Even with the right docs loaded, it can quote the wrong passage confidently. Spot-check answers and ask it which document it used.