When an AI project disappoints, the data underneath it is usually the reason — not the model, not the tool, not the vendor. The clever part of AI is now a commodity; the part that decides whether it works for you is the messy, unglamorous information sitting in your systems. And here's the good news most people miss: you don't need perfect data. You need data that's fit for the specific job you're asking AI to do.
That distinction matters, because "get your data in order" is the kind of advice that sounds sensible and stops everything dead. It conjures a multi-year programme, a new platform, a small fortune. In practice, getting ready for one AI use case is far smaller and far more doable than that.
What "ready" actually means
Strip away the jargon and ready data comes down to four plain qualities. It's accessible — you can actually get to it without a week of exporting and stitching. It's reasonably accurate — close enough to reality for the decision at hand. It's consistent — a customer is the same customer across your systems, not three slightly different spellings. And you're allowed to use it — the privacy and consent basics are in order.
Notice what isn't on that list. You don't need a data lake, a warehouse, or a "single platform" before you start. Those can come later, if a use case ever justifies them. The myth that you must build the cathedral before you light a single candle is exactly what keeps sensible firms on the sidelines.
The four questions
Before you ask whether AI can help, ask these four things about the data it would rely on. They're the same four qualities above, turned into questions you can actually answer.
1. Can you get to it?
If the information lives somewhere you can reach — a system with an export, an API, a clean spreadsheet — you're in business. If getting it means three people, a fortnight, and a lot of copy-paste, that's the first thing to fix.
2. Is it accurate enough?
Not perfect — enough. If half your contact records are years out of date, an AI that drafts outreach will simply be wrong faster. Know roughly how good the data is before you trust a machine to act on it.
3. Is it consistent across systems?
The same customer, product, or order should mean the same thing wherever it appears. When your CRM, your accounts package, and your spreadsheets each tell a slightly different story, AI inherits the confusion.
4. Are you permitted to use it?
Personal data carries obligations. Before feeding anything into a tool, be clear on what you hold, why, and whether your consent and privacy basics cover the use you have in mind.
You don't need to fix everything
The biggest trap is scope. Faced with imperfect data, the temptation is to launch a "data quality programme" that tries to clean everything, everywhere, before anything useful happens. That's how budgets vanish and momentum dies.
Scope the data work to the one use case in front of you. If you want AI to triage inbound enquiries, you need the enquiry data in good order — not your entire estate. Pick the use case, ready the slice of data it depends on, prove the value, then move to the next. Boil one kettle, not the ocean.
Common traps
Most data trouble isn't dramatic. It's the slow, ordinary friction that builds up in any growing business. The usual suspects:
Data trapped in silos. The information AI needs is split across a CRM, an accounts tool, a couple of apps and a folder of spreadsheets — none of which talk to each other. Siloed data is consistently named the single biggest obstacle to readiness.
Re-keying. If someone is manually copying numbers from one system into another, you have both a productivity leak and a quality problem — every re-key is a chance to introduce an error AI will then dutifully repeat.
No single source of truth. When nobody can say which system is authoritative for a given fact, every report becomes a debate. AI can't resolve a disagreement your own systems are having.
A readiness check
You don't need a consultant in the room to get a feel for where you stand. Run through these for the use case you have in mind:
- Can you name the one system that holds the data this use case needs — and get an export of it today?
- Do you have a rough sense of how accurate and up to date that data is?
- Does a key record (customer, product, order) mean the same thing across the systems involved?
- Is anyone re-keying this data by hand between tools right now?
- If it includes personal data, are you clear on your consent and privacy position?
- Have you scoped the data work to this one use case — not to "all our data"?
If most of these are a confident yes, you're readier than you think, and the right next step is a use case, not a data programme. If they're mostly no, that's not a reason to wait — it's simply the work to do first, on a slice small enough to be worth doing.
Wondering if your data's ready for AI?
We start with the use case, not a year-long clean-up — and we'll tell you honestly whether the data behind it is fit for the job, or what to fix first.
Book a call Start a conversationSources: Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk" (February 2025); PEX Report 2025/26, reported via AI Data & Analytics Network (2025); Informatica 2025 CDO Insights survey (2025). Figures vary by survey definition and sample; treat as indicative of direction, not precise benchmarks.