How to give AI your business data so it writes accurately
How AI uses your own documents, what to prepare, what to leave out, and how to organize it for better, cheaper results.
Ask a general AI model for your prices and it will either say it doesn't know, or make up something plausible. To write accurately about your business, it needs your information. Here's how that works and how to do it well.
How AI uses your documents
You don't have to train a model to use your own data. Most tools use a simple approach, usually called retrieval:
- Your documents are split into short passages.
- When you make a request, the system searches for the passages that match it.
- Those passages are sent to the AI along with your request.
- The AI writes using them.
So the AI isn't memorizing your business. It's being handed the relevant pages each time, like a colleague given the right folder. This matters for two reasons:
- Quality depends on what gets found. A messy library means the wrong passages get found.
- Cost depends on what gets sent. AI models charge by the amount of text. Sending only what's relevant is far cheaper than pasting in everything.
What to add first
Begin with what you keep correcting in drafts. For most businesses, that's:
- What you sell: products, services, prices, what's included.
- Who it's for: your customers, and their real questions and objections.
- How you work: delivery areas, opening hours, policies, guarantees.
- Proof: results, case studies and customer stories that you have permission to use.
- Past writing: your best posts and emails, kept separate from reference facts.
What makes data easy to use
Search works best on clear, plain text. A few habits help a lot:
One topic per document. "Refund policy" is easy to find. "Company handbook, 80 pages" is hard to find anything in.
Use the words customers use. If people ask about "delivery" and your document only says "fulfilment", a search can miss it.
Keep it current. Old prices cause more damage than missing ones. When something changes, replace the document.
Remove repeats. If three documents give different prices, the AI may pick any of them.
Check scanned files. A scanned PDF is an image, so it may contain no readable text. Run text recognition on it first.
What to leave out
Everything you add may be sent to an AI provider when it's relevant, so think before uploading:
- Passwords, API keys and financial account numbers.
- Health records, ID numbers and other sensitive personal details.
- Other people's personal data, unless you have a right to use it.
- Confidential client material, unless your agreements allow it.
Read your tool's privacy policy to see where your data goes. Here's ours.
Organize by job
One big pile of data makes every search compete with everything else. Group it by what it's for instead:
| Group | Contains | Used by |
|---|---|---|
| Menu and prices | Products, price list | Posts, emails, ads |
| Brand voice | Your story, best posts | All marketing |
| Wholesale | Terms, delivery areas | Proposals, contracts |
Now your contract writer never pulls from your Instagram captions, and your social writer never quotes clause 12. Results improve and costs fall, for the same reason.
Test before you trust it
Before relying on it, ask a question you know the answer to, and check which passages come back. If the right one isn't found, the usual fix is to reword the document, or to split it up. If the wrong one comes first, look for an old version to remove.
Even with good data, read the draft. Retrieval reduces mistakes, but it doesn't eliminate them, and a model can still misread a passage.
How this works in DeepRit
In DeepRit, you add files, web pages and notes to data groups, and each writer imports only the groups it needs. Search runs in your browser and costs nothing, and Test your data shows which passages a request would find. Only the matching passages are sent with a request. See the whole system.