Most AI projects that disappoint don’t fail because the model wasn’t clever enough. They fail because the information it was given was incomplete, out of date, contradictory, or not something the business was allowed to use that way.
The useful part is that you can check this before you spend anything on AI. Data readiness isn’t a mystery. It comes down to four questions: do you know where your data is, do you trust it, can the right people (and only the right people) reach it, and are you allowed to use it for this?
The checklist
Work through these ten questions for the data your AI idea would depend on. Be honest. A “sort of” counts as a no.
Where it lives
1. Can you list the systems that hold your core data? Accounting, inventory, job management, payroll, CRM, email and shared drives, plus the spreadsheets that sit between them. The spreadsheets matter most, because they’re usually where the real process lives.
2. Is there one agreed source for each key fact? If your customer list exists in your CRM and your accounting system, and they disagree, which one wins? If nobody can answer quickly, an AI tool won’t be able to either.
Whether you can trust it
3. Are customers, products and suppliers recorded consistently? “ABC Ltd” and “A.B.C. Limited” as two separate customers. Product codes typed freehand. Suppliers set up twice. These look minor until something tries to add them up.
4. Is it current? When was the price list last updated? Are inactive customers and discontinued products marked as such, or do they still look live?
5. Do your reports reconcile? If the sales report and the accounting system give different totals, anything built on either will be wrong some of the time. Reconciling is often where the real findings are. In one analysis for a distributor, simply joining up stock, sales and accounting data surfaced $850K of slow-moving and dead stock. That’s the kind of thing you want to find before an AI tool starts quoting from it.
Who can reach it
6. Do permissions match roles? Look for files shared with everyone, former staff who still have access, and shared logins. AI tools generally see what the person using them can see, so any oversharing becomes easier to stumble on.
7. Is sensitive information identified? Payroll, health and safety incidents, customers’ personal details, commercially sensitive pricing. You can’t protect information you haven’t identified.
Whether you’re allowed
8. Do you know what personal information you hold, and why you collected it? Under the Privacy Act 2020, the purpose you collected personal information for limits what you can use it for later. The information privacy principles are a short read and worth knowing.
9. Do your contracts allow it? Some supplier, customer and partner agreements restrict how you use or share their information, including with third-party processors such as AI providers.
10. Do you know where an AI tool would send the data and what the provider does with it? Check whether the provider keeps it, whether it’s used to train their models, and where it’s processed. The Office of the Privacy Commissioner’s guidance on generative AI says not to enter personal or confidential information unless the provider has confirmed it won’t be retained or disclosed, or to strip out anything that could identify people first.
How to read your score
Mostly yes. You’re in good shape. Start with a focused pilot on one use case, with a person reviewing the output.
A mix. Fix the gaps for the use case you care about most, then pilot. Don’t try to fix everything first.
Mostly no. AI isn’t your next project. Getting your core data joined up, consistent and reconciled is, and it pays off on its own through better reporting, faster month-ends and fewer arguments about whose number is right. AI becomes much easier afterwards.
Fix it for one use case, not the whole company
“Data readiness” can easily turn into an endless programme. Avoid that by working backwards from a single, specific use case.
Say you want an assistant that can answer “what’s our margin on this product line this month?”. That doesn’t need a tidy SharePoint. It needs clean product, sales and cost data from your accounting and inventory systems, joined in one model, with clear rules about which figures are final. That’s a defined, finishable piece of work.
The steps:
- Pick one use case that would save real time or money if it worked.
- List the data it needs and the system each piece comes from.
- Run the ten questions for that data only.
- Fix the gaps you find, starting with consistency and reconciliation.
- Build a governed connection or model so the AI reads from one trusted place, with access controls that match your roles.
- Pilot with human review, measure whether it helps, then decide whether to expand.
Each use case you complete leaves your data in better shape for the next one. Over time, readiness takes care of itself.
Where to start
Pick the AI idea your team is most excited about and run the checklist against the data it would need. If the answers are mostly no, that’s useful to know now rather than after a failed pilot. Our data platforms work gets core data joined up and trustworthy, and our AI and agents work puts it to use safely. Get in touch if you’d like a second opinion on where you stand.