1. Can you name the questions?
Start with the questions people will actually ask, not the technology. Write down ten real questions from managers, such as “Which regions missed target last month?” or “Which customers are overdue by more than 60 days?”. If you can’t list them, the first step is a short discovery exercise, not a platform.
For each question, note who asks it, how often, and how they get the answer today. This becomes your test set later.
2. Do your key terms have one definition?
Ask three people how revenue, an active customer or a qualified lead is calculated. If you get different answers, an AI assistant will too. It can only be as consistent as the definitions it is given.
- Each core metric has a written definition and an owner.
- The definition is implemented once, in a shared model, rather than in individual spreadsheets.
- Reports and AI tools read from that same model.
3. Is the data joined up?
Many useful questions cross systems: sales from the CRM, invoices from finance, orders from the ERP. Check whether those systems share keys, such as customer IDs and product codes, and whether someone already reconciles them. If people match records by hand today, a data engineering step usually comes before AI.
4. Is it fresh enough for the decision?
Daily decisions need daily data. Monthly reporting may be fine with a monthly refresh. Agree the freshness each question needs, then check that loads run on schedule and that failures alert someone.
5. Do you trust the quality?
You don’t need perfect data, but you do need to know where it is weak. Look for duplicates, missing values in key fields, and records that change without explanation. Simple quality rules that run on every load, with an issue list someone owns, go a long way.
6. Are permissions in place?
An AI tool should never show people data they couldn’t see in a report. Check that access rules, such as row-level security by region or business unit, exist in the data model itself, so any tool built on it inherits them.
7. Can every answer be checked?
Decide up front how people will verify an answer. Good data agents show the query, filters and source behind every response. Before launch, test them against your list of questions with known correct answers, and agree an acceptable level of accuracy with the business owner.
What to do with the results
If you can tick most of this list, a focused data agent or AI workflow pilot is a reasonable next step. If several items are open, start with the foundation: agree definitions, join the key systems and put quality checks and permissions in place. That work pays off in better reporting even before any AI is added.