How do I make AI work with my own data?
The tool talks in generalities — can it learn my business? It can, and doing so does not require training a model. 📚
What most businesses need is not to “train” AI but to give it the right material: your own documents, your own language, your own rules.
Short answer: there are three tiers — adding to the brief, an information file, a connected source. Most businesses stop at the second tier, and that is enough. 🪜
The three tiers
BU BÖLÜMÜN ÖZETİ
- Tier 1: adding to the brief
- Tier 2: an information file
- Tier 3: a connected source
Cost and effort rise from the bottom up. 🪜
Tier 1: adding to the brief
Pasting the needed information into the brief each time. Free and fast; tiring on repetitive work. 📋
Tier 2: an information file
Products, pricing logic, frequent questions and tone rules in one file. Uploaded to the tool and used as the same base every time. 📄
Tier 3: a connected source
Company documents connected to a repository the tool answers from. This requires setup and maintenance. 🔗
What goes in the information file?
BU BÖLÜMÜN ÖZETİ
- What you sell
- Who you sell to
- How you speak
- What you never say
Between half a page and two; more does harm. 📄
What you sell
Short definitions of services and products, with scope limits. This is where the tool gets confused most. 🛍️
Who you sell to
Customer profile, sector, the person deciding. This sets the tone correctly from the start. 👥
How you speak
Words you use and words you avoid, tone rules. Brand language is defined here. ✍️
What you never say
Guarantee phrases, hype, competitor names. A forbidden list shortens correction time. 🚫
Which documents can be given?
The boundary is drawn by data type. 🔐
Can be given
Your own product documents, frequently asked questions, published content, anonymised case summaries. ✅
Never given
Identity and contact details, health data, signed contracts, personnel files, customer lists; the frame sits in the data security guide. 🚫
When is a connected source built?
BU BÖLÜMÜN ÖZETİ
- Condition 1: document volume
- Condition 2: frequent querying
- Condition 3: keeping it current
Three conditions sought together. 🔗
Condition 1: document volume
Too many documents to paste by hand. With few documents this setup is needless overhead. 📚
Condition 2: frequent querying
If the team looks up the same information several times a day, the setup pays for itself. 🔎
Condition 3: keeping it current
If documents are not updated, the system will confidently state outdated information — the most dangerous state. 📅
Three common mistakes
BU BÖLÜMÜN ÖZETİ
- Mistake 1: uploading everything
- Mistake 2: not updating
- Mistake 3: dropping verification
- The sentence that cuts all three
All three spoil the output. 🚧
Mistake 1: uploading everything
With conflicting and outdated documents mixed in, answers become inconsistent. Fewer, correct documents work better. 🗂️
Mistake 2: not updating
If a repriced product stays old in the file, the tool states the wrong thing confidently. 📅
Mistake 3: dropping verification
Working with your own data raises accuracy but does not guarantee it; checking sits in the verification guide. 👁️
The sentence that cuts all three
This: “Few, correct, current documents — and still check.” ✅
What should I do today?
BU BÖLÜMÜN ÖZETİ
- Step 1: write the information file
- Step 2: choose three documents
- Step 3: name an update owner
- If you want help
Three steps, one day. 🪜
Step 1: write the information file
Four headings: what, to whom, how, what we never say. Half a page is enough. 📄
Step 2: choose three documents
The three most useful ones. More is not uploaded at the start. 🗂️
Step 3: name an update owner
Who reviews it quarterly? An un-updated file becomes the source of the error. 👤
If you want help
Let us build your information file and pick the tier: use the consult your expert form. For your current usage see the business AI usage audit; the whole sits on the AI consultancy page. 🎯
Related reading from the archive: a setup with your own data · what you can share.
📝 Notes From the Field
A firm kept saying “the tool doesn’t understand our business”. A half-page information file was written: what was sold, to whom, in what language and which phrases were never used. The file was attached to the start of every job. The same tool and the same team began working with a third of the correction time.
📖 Short Glossary
Information file: a fixed text summarising product, audience, language and forbidden phrases. Connected source: company documents linked to a repository the tool answers from. Anonymisation: removing names and contact details. Update owner: the person refreshing the file periodically.
⚡ Quick Summary
Working with your own data does not require training a model. 📚 There are three tiers and most businesses stop at the information file. The file has four headings and stays under two pages. The boundary is drawn by data type; documents are kept few, correct and current.
🎯 Next Step
Let us build your information file and pick the tier: use the consult your expert form. The brief side sits in the brief guide; for your current usage see the usage audit.
Frequently Asked Questions
Sık Sorulan Sorular
With few documents and slow change, tier 2 is enough. With hundreds of documents and frequent updates, tier 3 is worth considering. ⚖️
Remove names, company names and contact details; the structure and content that remain are enough. 🧼
In the company account with the data setting off; the structure sits in the company account guide. 🏢
The technical side builds it, someone who knows the work prepares the content. Technical setup alone gets filled with the wrong documents. 👥
For most businesses no, and it is expensive. An information file and, where needed, a connected source deliver the same result at far lower cost and maintenance.
Between half a page and two pages is ideal. As it grows the tool starts confusing priorities; information requiring detail should be supplied as a separate document.
On business plans, keeping inputs out of training is usually the default and that setting should be checked. Even so, identity and customer data are never uploaded under any setting.