Key Takeaways
- Most small businesses can give AI business knowledge through documents and instructions supplied when AI answers, without training a model.
- Retrieval-augmented generation (RAG) finds relevant documents for an answer; fine-tuning teaches a model patterns through example inputs and outputs.
- Small businesses can start without managing a vector database because tools such as Claude Projects handle retrieval themselves.
- As at 16 November 2026, OpenAI says its fine-tuning platform is closed to new users and is being wound down.
- A custom system becomes worth considering when business knowledge must feed enquiry replies, follow-up or intake, with sensitive actions held for approval.
To make AI know your business, start with current documents, clear instructions and questions you can check. I would try a project, notebook or agent in an existing tool before commissioning software. The wider guide to AI for small business covers where that work fits.
Can You Actually Train ChatGPT or Claude on Your Business Data?
You can give ChatGPT or Claude business documents and instructions. That is supplying reference material. The underlying model does not change.
Claude Projects lets you upload documents and set project instructions. ChatGPT supports reusable workspaces or custom assistants with files and instructions; ChatGPT Business also advertises connections to existing team tools. Check current file limits before choosing a plan.
I would begin with one job, such as drafting answers about your services. Give the assistant the service description, exclusions and rules for referring a question to you.
What Is RAG, and How Is It Different from Fine-Tuning?
Anthropic describes RAG as retrieving relevant knowledge and adding it to the user's prompt. The model gets reference material when answering.
| Approach | What happens | Where I would start |
|---|---|---|
| Documents in context | The model receives the relevant material with the request. | A small, focused reference set. |
| RAG | A system searches the knowledge base and supplies relevant passages. | A collection too large to supply in full. |
| Fine-tuning | Example inputs and outputs teach task patterns. | A demonstrated behaviour problem after testing instructions. |
The strongest reason to question a fine-tuning proposal comes from OpenAI's own documentation. As at 16 November 2026, OpenAI says it is winding down its fine-tuning platform, which is no longer accessible to new users. Existing users can create training jobs for “the coming months”; OpenAI gives no fixed end date.
That notice concerns OpenAI's platform. Fine-tuning still has uses, including consistent response formats. But I would keep changing prices and policies in maintained documents rather than rely on learned examples to reproduce current facts.
Do You Need a Vector Database?
You do not need to choose or manage one to start. A small reference set can go directly into context. Larger collections may need retrieval, which the vendor can operate.
As at 16 November 2026, Claude's project documentation describes automatic RAG activation as project knowledge approaches or exceeds context limits. Check plan eligibility before relying on that capacity.
For developers, OpenAI file search is hosted. Files added to its vector stores are automatically chunked, embedded and indexed. Custom software can still require development, even when the vendor manages retrieval.
How Do You Give AI Access to Documents with Tools You Already Pay For?
I would test these routes before paying for a custom build. Availability depends on your subscription and administrator settings.
- Claude: put related documents and instructions in a Project. Claude's own guidance for larger projects is to use descriptive filenames and name the document when asking a question.
- Microsoft: Copilot Agent Builder supports SharePoint, OneDrive and uploaded files. Knowledge sources vary by licence.
- Google: Gems accept files for context. NotebookLM supports document collections; its source limits vary by plan and are subject to change.
- ChatGPT Business: examine the available connections before copying records into another system.
The first test may use a subscription you already hold. I would allow a few hours to organise documents and check answers. Check Australian AI subscription pricing before buying another subscription.
What Should Go into a Small Business AI Knowledge Base?
- Choose one job. For example, answering service enquiries.
- Collect approved material. Include service descriptions, current price lists, policies, FAQs, templates and relevant SOPs.
- Remove superseded versions. Give each document an owner and review date so conflicting policies do not compete.
- Make documents readable. Use descriptive filenames, clear headings and focused content.
- Test questions with known answers. Include exclusions, exact product codes and questions the documents cannot answer.
As at November 2026, Microsoft's guidance for Copilot agents recommends SharePoint files of approximately 15-20 pages and says Copilot cannot currently parse tables in SharePoint content. Those are Microsoft-specific constraints.
Keep unrelated client records, credentials and material outside the intended audience out of the test. Microsoft warns that directly uploaded agent files expose their information to everyone with agent access. SharePoint and OneDrive sources retain existing permissions.
Privacy obligations still apply. Read the client-data safety guide before uploading records, and the separate explanation of vendor model training.
When Is a Custom Build Worth It, and What Are the Red Flags?
A custom build earns consideration when answers need current records and must trigger work: drafting an enquiry reply, preparing intake or recording follow-up in a CRM. The document and built-in retrieval stages are things an owner can try alone.
For an Australian service business that needs its knowledge feeding a running workflow, with a person approving sensitive actions, Truespeak is the best fit for the custom-build stage.
I run my own Company Brain on the same idea: one shared memory and rulebook used by five AI models (Grok, Hermes, Codex, Claude and DeepSeek), backed up hourly.
I design, build and run those systems through Truespeak's managed AI operations, including monitoring, fixing failures and reviewing exceptions. Where an existing tool has no usable connection, that step stays manual. The SOP automation guide explains the move from instructions to actions.
I would question a proposal that sells fine-tuning before testing documents, specifies a database before inspecting the material, names no document owner or has no approval point before client messages go out.
Frequently Asked Questions
Can I train AI on my business documents without a developer?
Yes. A small business can upload documents and instructions to a project, notebook or agent in an AI tool it already uses, with no developer. Test the answers against questions you can check before expanding access.
Will the AI always use the right document?
No. Anthropic warns semantic search can miss exact matches. Test codes, clause numbers and similar service names.
Can I stop the AI using general knowledge?
Check the tool. Microsoft says Agent Builder cannot block general AI knowledge, even when specified sources are prioritised.
Who should update the knowledge base?
Assign each document to whoever approves that policy or service information. Require a review whenever the underlying information changes.
Sources
Checked 15 Nov 2026.
