What Data Do You Need Before Starting an AI Project?
Before starting an AI project you need four things: the documents and records the AI will read, real examples of the task done well, access to the systems it must look things up in, and a clear rule for who is allowed to see what. You rarely need a large dataset for training, because a typical business AI project uses an existing model and hands it your own content at the moment a question is asked. If personal data is involved, you also need to know what your privacy notice and contracts permit.
That moves the preparation from data science to housekeeping: finding the current version of each document, naming who owns it, and collecting examples. Whoever builds your custom software or AI feature will ask for these things early, so having them ready shortens the project.
Which documents will the AI read, and what condition do they need to be in?
The AI reads the material you would give a new employee: policies, product information, price lists, procedures, help articles and past tickets. Those documents need to be current, owned and consistent, because the model cannot tell which of two files is out of date and will answer confidently from whichever one it is given.
Current
Archive superseded versions before anything is connected. An old returns policy sitting next to the new one is a typical source of wrong answers.
Owned
Each document needs a named person who keeps it correct, and you need the right to use it. Material licensed from a third party, or supplied by a customer under contract, may not be yours to put into an AI tool.
Not contradictory
Where two sources disagree, decide which one wins and write that decision down. Also check format: text locked inside scanned images or heavily merged spreadsheets usually needs converting first.
Why do you need real examples of the task done well?
You need them because they are the only objective way to tell whether the AI is good enough. Collect real inputs, such as customer questions or incoming invoices, each paired with the output your best person produced. This becomes the test set that is run again after every change, so nobody has to judge quality by impression.
Anthropic's guidance on building evaluations recommends tests that mirror the real-world mix of tasks and include edge cases. In practice that means adding the awkward ones on purpose: requests with missing information, ambiguous wording, and questions the AI should decline or pass to a person.
Which systems does the AI need access to, and who is allowed to see what?
List every system the AI must look things up in, such as the CRM, order database, ticketing tool and shared drives. For each one, note whether it has an API or an export, who administers it, and whether the AI only reads or also writes. If it will take actions rather than only answer, you are planning an agent; see what an AI agent is.
Fix permissions before you connect anything
An AI assistant makes existing permissions matter more, because it finds in seconds what used to take deliberate searching. Microsoft's documentation, checked on October 6, 2026, says Copilot only surfaces organizational data to which individual users have at least view permissions, and tells administrators to rely on the permission models in services such as SharePoint. So a salary folder shared with everyone by accident becomes answerable by question. Decide who may see what, then correct the sharing settings to match.
Can you use customer or employee personal data in an AI project?
Often yes, but only within what you have told people and what your contracts allow. In the UK, the ICO's guidance on purpose limitation (last updated March 23, 2026) says you must specify your purposes in your privacy information and may only reuse personal information for a new purpose if it is compatible with the original one. In the US, the FTC stated in January 2024 that AI model providers who fail to abide by their privacy commitments to users and customers may be liable under the laws it enforces.
Three documents are worth reading before the project starts: your own privacy notice, your customer contracts (confidentiality and data location clauses), and the AI vendor's terms on training and retention. Leave out any personal fields the task does not need. This is not legal advice.
How much data is enough to start?
Enough means your documents cover the questions people actually ask and your examples cover the kinds of cases that actually occur. Coverage matters, not volume, and there is no universal number.
A practical test: take last month's real requests and check, for each, whether a correct answer exists in a document you would hand over. Every gap is something to write down before the project, not something the AI will work out. For examples, make sure each category of request appears several times, including rare cases that are expensive to get wrong. The same Anthropic guidance favors a larger set of automatically graded test cases over a small hand-graded one. If coverage is patchy, narrow the task until it is complete.
What are the common mistakes with AI project data?
The most damaging mistake is treating this as a volume problem and connecting everything. The others follow from skipping the housekeeping:
- Connecting a whole shared drive, drafts and duplicates included.
- Testing only with easy examples, so failures appear in front of customers.
- Discovering over-shared folders after launch instead of before.
- Assuming knowledge held in people's heads is written down somewhere.
- Naming no owner for the content, so answers decay as the business changes.
- Sending personal data to a vendor without reading its terms.
What should you do next to get ready?
Pick one task and work through this readiness checklist. Unticked items are not reasons to cancel; they are the first tasks in the plan.
- Write the task in one sentence and name the person who will judge the output.
- List the documents the AI will read, archive old versions, and name an owner for each.
- Resolve contradictions between sources and record which one wins.
- Collect real examples with the correct output, including hard cases.
- List the systems involved, how each can be accessed, and who administers it.
- Review sharing settings so access matches who should see what.
- Identify any personal data and check your privacy notice, contracts and vendor terms.
Quick answers
Do I need to train my own AI model for a business project?
Usually not. A typical business AI project uses an existing model and gives it your documents and records when a question is asked, so the preparation is about content and examples, not training data.
How many documents do I need for an AI project?
There is no fixed number. You have enough when the documents contain a correct, current answer to the questions people really ask about the task you have chosen.
What is a test set in an AI project?
A test set is a collection of real inputs paired with the correct output, used to check the AI after every change. It should include difficult and ambiguous cases, not only easy ones.
Can an AI assistant show staff files they should not see?
It can if those files are already shared too widely or the AI is connected through an account with broad access. Review sharing settings and make the AI follow each user's own permissions before launch.
Can I put customer personal data into an AI tool?
Only if that use fits what your privacy notice says, what your customer contracts allow, and the AI vendor's terms on training and retention. Check all three first and take legal advice where unsure.
Conclusion
The data you need before starting an AI project is mostly material you already have: current documents with an owner, real examples of good work, a list of systems, and a decision on who sees what. Getting those in order is inexpensive compared with building on top of contradictory content, and it tells you early whether the task you picked is realistic.
Entrant Technologies builds websites, web applications, mobile apps and custom software. If you would like a second opinion on whether your content and systems are ready, contact us and we will take a look.