To prepare your business data for AI, work through six steps: list the data you have, give each source an owner, clean it, tighten who can access it, decide how personal information is handled, and keep one approved version of your FAQs and policies. Do this for the data your first use case needs, not for everything.
Do you need a full data strategy before using AI?
No. Microsoft’s AI adoption guidance says AI does not require an organization to create an entirely new data strategy before adoption begins. The data work depends on the use cases you pick. The instruction is simple: identify the data each AI capability requires, and make sure you can access it securely and govern it appropriately.
That keeps the job small. A chatbot that answers service questions needs your FAQs, policies, hours and pricing rules. It does not need ten years of accounting records.
The six-step data preparation checklist
1. Take an inventory
List what you have before you connect anything. Microsoft’s planning guidance recommends documenting data sources, formats, quality and accessibility across the organization, and notes that this inventory reveals how much preparation each use case needs.
The Federal Trade Commission gives the same starting point for personal information. The first principle in its guide for businesses is to take stock: know what personal information you have in your files and on your computers. The guide suggests tracing how data enters the business, where it ends up and who has access to it.
A simple spreadsheet is enough. For each source, record:
| Column | Example |
|---|---|
| Source | Shared drive folder “Policies” |
| Format | Word documents and scanned PDFs |
| Owner | Operations manager |
| Contains personal data? | No |
| Last reviewed | March 2026 |
| Needed for first AI use case? | Yes |
2. Assign an owner to every source
Data without an owner goes stale. The NIST AI Risk Management Framework core puts documented roles, responsibilities and lines of communication under its Govern function, and that idea applies directly to content. Someone specific should be able to answer two questions about each document: is this still correct, and when was it last checked?
Ownership does not need a new job title. It needs a name next to each row in the inventory.
3. Clean what the AI will read
Microsoft’s planning guidance says data quality directly affects AI performance and determines which use cases you can implement successfully. For most small and mid-sized businesses, cleaning means four things:
- Remove duplicates and old versions. If three versions of a refund policy exist, an AI system may retrieve any of them.
- Resolve conflicts. When the website says one thing and the internal handbook says another, pick the right answer and correct the other.
- Make files readable. Scanned images of documents need text extraction before most AI tools can use them.
- Fill the gaps. Write down the answers that currently live only in someone’s head.
4. Fix access permissions before launch
This step is easy to overlook. An AI assistant that works inside your business systems typically follows the permissions already in place. Microsoft states that its Copilot only surfaces organizational data to which individual users have at least view permissions, and says it is important to use permission models so the right users or groups have the right access to the right content.
Read that from the other direction. If a folder of contracts or payroll files is shared with everyone, an assistant can bring that content into an answer for anyone who asks.
The standard to aim for is least privilege. The NIST glossary definition describes it as granting each entity the minimum system resources and authorizations it needs to perform its function. The FTC guide uses the same principle: each employee should have access only to the resources needed to do their job.
Practical checks:
- Review who can open each folder in the inventory.
- Remove “anyone with the link” and company-wide sharing where it is not needed.
- End access when employees leave or change roles.
- Give a customer-facing chatbot access only to content you would be comfortable publishing.
5. Decide how personal data is handled
The FTC’s five principles are a workable policy for small businesses: take stock, scale down, lock it, pitch it, and plan ahead. Two of them matter most for AI.
Scale down. The FTC advises keeping only what you need for your business, and not collecting or keeping sensitive personal information without a legitimate business need. Applied to AI: if the use case works without names, account numbers or health details, leave them out of the data the system can read.
Lock it. Protect the information you keep. Before you connect a tool, read the vendor’s documentation on how data is stored, who can see it and whether it is used for training.
For a broader structure, NIST publishes a Privacy Framework, a voluntary tool meant to help organizations identify and manage privacy risk. The NIST AI Risk Management Framework also asks organizations to map legal and other risks that come from third-party software and data, which covers the AI vendors you choose.
This section is general information, not legal advice. Privacy rules vary by state, industry and contract, so confirm your obligations with your own counsel.
6. Build a single source of truth for FAQs and policies
An AI system that answers questions from your documents needs one approved version of each answer. For every FAQ and policy, keep:
- One document, in one place.
- A named owner and a “last reviewed” date.
- Plain wording that matches how customers ask.
- Short sections that make sense on their own.
Then keep it current. The AWS overview of retrieval-augmented generation recommends updating both the documents and their indexed versions, in real time or in periodic batches, so answers reflect the latest content. If you are new to how chatbots read documents, what is RAG explains the mechanics.
How do you know your data is ready?
Use these checks before you sign off:
- Every source needed for the first use case appears in the inventory with an owner.
- Each FAQ and policy has exactly one current version.
- A person outside the team can find the right answer to ten common questions using only those documents.
- Sharing settings have been reviewed, and nothing sensitive is open to everyone.
- You can state in one sentence what personal data the AI system can read, and why.
If any check fails, fix that item first. Correcting a document before launch is simpler than correcting an answer a customer has already received. Data readiness is one part of a wider set of questions covered in AI readiness: 10 questions to ask before you invest, and the risk side is covered in AI risk management for small business.
The short version
Prepare the data your first AI use case needs, not all of it. Inventory it, name an owner, clean it, tighten permissions, minimize personal data and keep one approved version of every FAQ and policy. An AI system repeats what it can read, so accurate and well-governed content is the foundation of accurate answers.



