How to prepare your business data before an AI project

The short version

An AI project can get stuck when the information it needs is scattered, out of date, or known by only one person. A little preparation before you start makes the project smoother and the results more useful.

You do not need a data team. You need to know what you have, where it lives, and what is safe to share.

What we mean by business data

Data is not only numbers in a database. For most small businesses it means the everyday material your team already uses:

• Price lists, menus, and product descriptions

• Answers you give customers again and again

• Spreadsheets of orders, bookings, or customers

• Documents that explain how you do things, such as steps and checklists

• Chat and email history, when it is appropriate to use it

Step 1: list what you have and where it lives

Write a simple list: what information exists, who keeps it, and where it is stored. Include paper notes and personal phones, because that is often where the real answers are.

The goal is a map, not a perfect archive. Spend an afternoon on it, not a month.

Step 2: choose the information for the first task

Do not prepare everything. Pick the one task you want AI to help with, and gather only the information that task needs. If the task is answering customer questions, that means your common questions, your approved answers, and your current prices.

Step 3: clean up the basics

Remove duplicates, fix obvious mistakes, and update anything that is out of date. If two documents disagree, decide which one is right. AI will repeat whatever it is given, including old prices and conflicting rules.

Use clear names and consistent wording. If you call the same product three different names, tell your team which one to use.

Step 4: decide what is private

Mark what must stay inside your business: customer personal details, payment information, passwords, and anything covered by a contract. Keep these out of tools you do not fully understand.

Personal data may be covered by data protection rules. If you are unsure what applies to you, ask a qualified adviser before you share it.

Step 5: write down what good looks like

Collect a handful of real examples of good results: a reply you were proud of, a summary that was accurate, a report that was clear. These become the standard to test against.

Also note a few bad examples and why they were bad. Both kinds help whoever builds or tunes the solution.

Step 6: name an owner

Information changes. Someone should be responsible for keeping it current, such as updating a price or retiring an old answer. Without an owner, even a good setup slowly drifts out of date.

Make a small reference pack

For a customer reply pilot, your pack could contain one current price list, a delivery guide, and a page of approved answers. Put a date and an owner on each file. Add a short note saying which file wins if two answers disagree. Keep a separate list of questions that need a person.

Use made-up names in your test cases when a real identity adds nothing to the task. For example, you can check whether a delivery reply is correct without sharing the full customer conversation. Ask yourself what information is needed for the answer, then leave the rest out of the reference pack.

Check it with someone who does the work

Ask a teammate to answer a few sample questions using only the pack. If they need to ask you for missing details, the AI will need those details too. Fix the gaps before the pilot. Keep a note of each change so you can explain why a result changed later.

What you can skip

You do not need to move everything into one system first. You do not need perfect formatting. You can start with a small, current set of information and add more once the first task works.

Where we can help

If you are not sure what to gather, an AI Discovery session can map it with you. If you already know the task, our AI Build service starts by looking at the information your team has today. Book a free consultation and tell us what you have.