Is Your Data Ready for AI? The 3-Question Readiness Check

Plenty of AI ideas never reach a pilot. They stall on one sentence, "our data isn't ready", which sounds responsible and almost never gets tested. Here's a test you can run on one use case this week: three questions in order, the few blockers that are genuinely real, and what to fix when an answer is no.

Why we wrote this

We hear some version of "our data's a mess" in most early conversations about AI, often from the person who'll be asked to fix it. It's offered as a reason to wait, and the wait quietly becomes the plan. Meanwhile the team keeps doing the work by hand, and nobody checks whether the data was good enough all along.

So we've written down the test we run ourselves. It's for project owners and operations leaders who've been handed the data problem and want a clear way to size it. If you already have an AI use case live in production, you've likely worked through these questions already.

When "our data's a mess" ends the conversation

The AI idea finally made the planning agenda. Claims triage: stop assessors reading every document from scratch. Everyone agreed it was worth doing.

Then someone said, "Before we touch AI, we need to sort out our data." Nobody argued. Everyone has seen the customer table with four spellings of the same company.

So the idea moved to next year, under a new line called "data foundations". You're the one who has to write that plan. And nobody can tell you what "sorted" actually means.

The short answer: You don't need all your data ready, only the data behind one use case. It's ready when you can get to it, it describes the decision you want help with, and you'd trust a capable person to make that decision using it. Fix only what fails those three tests. 

Why "fix the data first" rarely ends

"Fix the data first" sounds responsible. In practice it becomes a cleanup project with no finish line, because nobody has defined what the data is being fixed for. It competes with every other priority and loses.

Here's what most vendors won't tell you: messiness is rarely the real blocker. Current AI models read inconsistent names, free-text notes and odd formats far better than your reporting tools do. The gap we see more often is that nobody ever recorded the outcome, so there's no history of what a good decision looked like. 

Three questions that decide it

Pick the use case first, then ask these three questions about the data behind it. Go in order, because each one is pointless if the one before it fails. 

1. Can we get to it?

This is about access, both technical and permission. Data you can't extract or aren't allowed to use doesn't count, however good it is. 

• What it looks like: records locked in a vendor system with no export, PDFs scattered across shared drives, a critical spreadsheet on one person's laptop.

• What good looks like: you can name the system, pull a sample of the last few months within a week, and someone with authority signs off on using it.

• If the answer is no: ask for a read-only export or API access to the one source you need. Don't start a data lake to solve an access problem.

2. Does it describe the decision?

Data can be complete and still useless for AI. The question is whether it holds what a person looks at when they decide, and what they decided.

• What it looks like: in an illustrative insurance case, the claims system records the payout, but not why the assessor flagged the claim. The inputs are there; the reasoning isn't.

• What good looks like: for a sample of past cases, you can see what came in, what was decided, and ideally whether it turned out to be right.

• If the answer is no: add a simple reason field to the workflow and capture it for a few weeks. Or have your most experienced person label a sample of past cases. Both are cheaper than a cleanup.

3. Would we trust a person using it?

This is the quality bar, and it's lower than most teams think. The test isn't "is it perfect". It's "would we let a capable new hire make this decision using only this data". 

Three yes answers means the data is ready for that use case. It says nothing about the next one, and that's fine. 

When the data really isn't ready

Some blockers are real, and no clever model gets around them. Three are worth checking before you go further. 

01

Nothing has been captured

If the decision happens in people's heads or on phone calls, there's nothing to learn from. The honest first step is capturing it, which may take a quarter.

02

Consent or legal limits

Under the Privacy Act, personal information is generally used for the purpose it was collected for, and sensitive information needs particular care. 

03

Data residency in health or finance

If you're APRA-regulated, CPS 234 has required resilience against information security incidents since July 2019, and an AI service holding your data is part of that picture.

What readiness looked like for AWN

Australian Wool Network (AWN) shows why "is it clean?" is often the wrong question. Market data outpaced the valuation algorithm every weekend as the market shifted. The challenge was pace: new information arriving faster than the system could absorb it. 

What drives the effort

This isn't free, and the cost depends on your data more than on the AI. Four things move it most. 

Number of sources

One system with an export is quicker to work with than five that disagree with each other.

How unstructured it is

Tables are quicker than emails, PDFs and scanned forms, which need extracting before they're useful.

Sensitivity and storage

Personal, health or financial data adds access controls, audit trails and residency decisions. Each one takes design time.

Whether outcomes were recorded

If past decisions were never captured, someone has to label a sample or capture new ones. That's expert time, not engineering time.

How we approach it

Check your own data

Three yes answers and none of the blockers above means your data is closer to ready than "next year" suggests.

If you'd like a straight answer on whether your data can support your top use case, book a call. We'll tell you honestly whether it's worth pursuing, and what we'd drop.