Is your data ready for AI? Five questions for small teams

The AI question often arrives as a demo. Someone sees an assistant answer a question in seconds, and the next meeting asks: could we do that with our data?

It is a fair question. The honest answer depends less on the model than on the data underneath it. A model can only work with what it can reach, and it will give a fluent answer over messy records just as easily as over clean ones.

So before you buy a tool or hire anyone, including me, answer five questions. They work for a small nonprofit, a startup, or an IT team inside a school, and none of them needs a data scientist.

For each one you get why it matters for AI, a quick self-check, and one small step you can take this week.

1. Where does your data live?

AI tools do not go looking for your data. Something has to connect them to it, and every place your data lives is another connection to build, secure, and keep running.

If answering "how many people did we serve last quarter?" takes a CRM export, two spreadsheets, and someone's inbox, an assistant will hit the same wall a new hire would.

You are in decent shape if these are true:

• You can list every system that holds records about the people you serve, your donors, or your customers.
• When two systems disagree, you know which one is the source of truth.
• No important number depends on a file only one person can open.
• You know how data moves between systems: typed in, imported, synced, or copied by hand.

A first step this week: make a one-page data map. Write one line per system, with what it holds, who uses it, and how data gets in and out. A shared doc is enough.

2. Who owns it?

Ownership has two meanings here, and AI needs both.

The first is control. Is the data in an account your organization holds, under terms that let you export it and connect other tools? The second is accountability. When a field is wrong, who fixes it, and who decides what it means?

An AI project with no clear owner on either side can stall at the first permission request or the first disputed number.

You are in decent shape if these are true:

• Each system on your map has a named owner, not "everyone."
• Admin access belongs to the organization, not to a former volunteer's, contractor's, or employee's personal account.
• You know whether your plan with each vendor lets you export your data or connect to it through an API.
• Someone can say who decides what a word like "active" or "enrolled" means.

A first step this week: add an owner column to your data map. Where the honest answer is "nobody" or "someone who left," write that down. That gap is your first fix, and it costs nothing to name it.

3. How clean is it?

Clean does not mean perfect. It means consistent enough that the same question gets the same answer twice.

AI does not fix messy data. It reads it faster. Duplicate records become double counts. A free-text field with five spellings of the same program becomes five programs. And a model will not always stop to tell you the data looks odd.

You are in decent shape if these are true:

• Duplicate records are rare, and someone notices when they appear.
• Dates, names, and categories follow one format.
• Key fields are filled in as a habit, not only when someone remembers.
• Two people who pull the same report get the same number.

A first step this week: pick one list that matters, such as contacts, enrollments, or donations, and spot-check 20 rows. Count the duplicates, the blanks, and the values that do not match. You are not fixing anything yet. You are finding out how big the job is.

4. What question do you actually want answered?

"We want to use AI" is not a project. "Which programs had people sign up and not show up last term?" is.

A specific question tells you which data you need, how clean it has to be, and how you will know the answer is right. It also keeps the first version small, which is where a small team should start.

In an earlier note on asking LMS and SIS data in plain English, the first version covered a short list of questions people already asked every week. Everything else got a clear "I can't answer that yet," not a guess. The same rule works here.

You are in decent shape if these are true:

• You can write the question in one sentence, in the words your team already uses.
• Someone answers it today, by hand, on a regular basis.
• You know what you would do differently with the answer.
• Someone already knows roughly what the answer should be, so a wrong one would stand out.

A first step this week: collect the three questions that interrupt someone on your team most often. For each one, write down who asks it, how often, and where the answer comes from today. Then pick the one with the clearest answer, not the most impressive one.

5. What are the privacy and consent rules?

Every record about a person came with an expectation about how it would be used. A donor gave an address for receipts. A family shared a student's details for enrollment. A client told a caseworker something in confidence.

Putting that data in front of an AI tool is a new use. Some of it may be off limits no matter how useful it would be.

The tools matter too. Before anyone pastes records into an AI product, read its data terms: where the data goes, how long it is kept, and whether it can be used to train models.

Specific rules may also apply, depending on your work. Student records, for example, may fall under FERPA, and health information or data about children may carry its own requirements. This is a practical checklist, not legal advice, so take anything specific to your counsel.

You are in decent shape if these are true:

• You know which fields are sensitive, and they are marked or kept apart.
• Your privacy policy and consent forms say what the data can be used for, and someone has read them recently.
• Staff know which AI tools are approved, and which data never goes into a public chatbot.
• Access is limited to the people who need it.

A first step this week: write a one-paragraph AI data rule for your team. Say which tools are approved, which data can go into them, and who to ask when it is unclear. It will not be perfect, and it is still better than everyone guessing.

What your answers tell you

If you answered all five with confidence, you are likely ready for a narrow first version: one question, one or two data sources, a small group of people using it, and a clear "I can't answer that yet" for everything else.

If you stalled on one or two, that is not a failure. Those gaps are the real first project. Naming an owner or cleaning one list is less exciting than a chatbot, but it is the work that makes a chatbot worth trusting.

Where I can help

If you want a second set of eyes, that is what my technical consulting is for: a scoped review, agreed up front, with a written answer at the end. Here, that answer covers what is ready, what is not, and what to do first.

When the answer is to build something, AI consulting picks up from there. It starts with an audit of the data, the people who will use it, and the ways it could fail.

Bring your data map and your three questions to a 30-minute call. That is enough to tell whether you need a cleanup, a narrow first build, or nothing yet.