Contact Us
Contact Us
Data & Analytics

AI-Ready Data: 7 Signs Your Data Isn't Ready (and What to Fix First)

Seven signs your data will stall an AI project, and the order we fix them in, starting with ownership and ending with monitoring.

JRC Software Team
7 min read
AI-Ready Data: 7 Signs Your Data Isn't Ready (and What to Fix First)

The demo went well. The support assistant answered every test question, the board liked it, and the pilot got its budget. Three weeks into real use, it told a customer about a refund policy that had been replaced years earlier. Nobody had touched the model. The old policy was still sitting in the document library, and nobody owned the job of retiring it.

That's what a data problem looks like in an AI project. AI-ready data isn't perfect data. It's data that's fit for one specific use case: someone owns it, its meaning is written down, its quality is measured, and a pipeline keeps it current once the system is live.

It's far cheaper to find the gaps before you fund the build than after the pilot has been promised to the business. Below are the seven signs we look for, and the order we fix them in.

Gartner's warning, and a framework you already know

In a May 2025 article for Technology Decisions, Gartner's Roxane Edjlali wrote that Gartner predicts 60% of AI projects without AI-ready data will be abandoned by 2026. The same piece reports that 63% of organisations either lack, or aren't sure they have, the data management practices AI needs.

Two of its points match what we see on the ground. Readiness is judged against a use case, so the data has to represent the patterns, errors and outliers that use case will actually meet. And it's "not one and done": the article recommends DataOps and data observability so you can see when data changes underneath you.

None of this is new thinking. The Australian Bureau of Statistics' Data Quality Framework (now an archived page, but still a useful checklist) uses seven dimensions: institutional environment, relevance, timeliness, accuracy, coherence, interpretability and accessibility. Every one of them applies to AI. What changes is how failure shows up. A bad report looks wrong. A model fed stale or contradictory records produces fluent, confident answers that look right.

7 signs you don't have AI-ready data yet

Run this against the data behind your planned use case, not your whole estate. If three or more apply, sort out the foundations before anyone writes a prompt.

  1. Nobody can say who owns it. Ask who decides what "active customer" means, or who signs off a change to the product hierarchy. If the answer is "IT", nobody is accountable, and the defects your pilot uncovers will sit in a backlog forever.
  2. The same number has three answers. Finance, sales and the dashboard each report a different customer count. A model grounded on that data inherits the argument. It can't settle it for you.
  3. Critical facts live in free text or empty fields. Hand-typed categories, a notes field doing the work of five structured columns, an "optional" field that's mostly blank. Retrieval and classification both depend on consistent structure.
  4. The pipeline is a person. Someone exports a CSV every Monday and drops it in a shared folder. There's no freshness guarantee, no audit trail, and no way to run the model on current data in production.
  5. You don't know where the personal information is. Contact details, health notes or financial data spread across systems with no labels. That blocks responsible use with an external model and creates risk under the Australian Privacy Principles.
  6. There's no history. Records are updated in place, so you can't see what a value was last quarter. Churn, forecasting and anomaly detection all need change over time.
  7. Quality is checked by complaint. Errors surface when a customer or executive spots them. There are no automated checks on row counts, null rates, duplicates or value ranges, and no alert when a feed stops arriving.

The order we fix things in

A data readiness assessment should end with a short, ranked list of fixes for one use case. Not a two-year transformation roadmap. Each step below makes the next one cheaper.

OrderFixSigns it addressesFirst deliverable
1Pick one use case and name a data owner1, 2One-page brief: owner, users, success metric
2Agree definitions for the 10 to 20 terms the use case relies on2, 3A short glossary mapped to source columns
3Classify sensitive fields in scope5Field-level tags for personal and confidential data
4Replace manual exports with scheduled ingestion4, 6Automated loads into a governed store, history kept
5Add automated quality tests3, 7Completeness, uniqueness, range and freshness tests with alerts
6Monitor in production7A quality dashboard and a named person who acts on it

Ownership goes first because every later step needs someone with the authority to make a call. Pipelines come before model work because a model built on a one-off extract has to be rebuilt the day it goes live.

For history, the usual pattern is a type 2 slowly changing dimension: instead of overwriting a row, you close it off with a valid_to date and insert a new version. Most modern tooling, dbt snapshots included, will handle this once you tell it which column identifies the record and which columns to watch.

Data quality for AI: set thresholds you can test

Teams stall because they think data quality for AI means cleaning everything first. It doesn't. It means the fields the use case actually reads meet thresholds you've agreed and can test.

For a support assistant grounded on knowledge articles, the critical checks are that each article has an owner, a review date and a product tag, and that superseded versions are out of the index. For a churn model, it's complete transaction history, one stable customer identifier and a written definition of "churned".

Write thresholds as testable statements, for example "fewer than 2% of active customers missing a segment" or "the orders table loaded within the last 26 hours". Then run them on every load. In dbt that's a not_null, unique or accepted_values test plus a source freshness check. Great Expectations or Soda do the same job in other stacks. The tool matters much less than having the rule written down and wired to an alert someone reads.

Metadata carries more weight with AI than with dashboards. The Gartner article argues for moving from passive to active metadata, and for pipelines that serve both training datasets and live production feeds. In practice, each table needs a description, an owner, a freshness expectation and lineage back to its source. That's the same foundation we describe in our lightweight data governance framework.

Three traps we keep seeing

The pilot that works on a hand-picked sample, then falls over on the full dataset. Nobody measured the gap between the two. Profile the production data before the pilot starts, not after it fails.

Assuming documents are ready because they're "just text". Policies, contracts and wiki pages need the same basics as tables: an owner, a review date, a classification, and old versions removed. Retrieval will surface a 2019 policy just as confidently as the current one.

Skipping the success metric. Without an agreed measure of value, data work looks like overhead and gets cut first. We go into this in why most AI pilots never reach production.

How JRC Software gets your data AI-ready

Our data engineering, analytics and insights team starts with one use case. In week one we sit with the people who own the source systems, profile the tables and documents the use case depends on, and score them against the seven signs above. Profiling means real numbers: null rates, duplicate keys, how old the latest load is, how many spellings a "category" field has.

At the end of the assessment you get a ranked fix list with effort estimates, a draft glossary for the key terms, and a map of where personal information sits in scope. If you go ahead, we build the foundations for that use case: automated ingestion, a store that keeps history, quality tests on every load, and sensitive-data tags.

We'll also tell you what not to do. Usually that's "don't start with a company-wide data catalogue" and "don't fine-tune anything until retrieval over clean data has been tried". Our engineers in Melbourne and Sri Lanka work as one team, which keeps the work moving across both time zones.

If you're about to fund an AI initiative and want a straight read on your data first, book a data readiness review and we'll tell you what to fix first.

Frequently asked questions

What is AI-ready data?

AI-ready data is data that's fit for one specific AI use case. It has a named owner, agreed definitions, measured quality, proper handling of personal information and an automated pipeline that keeps it current in production. It doesn't need to be perfect across the organisation. It needs to meet agreed, tested thresholds for the fields the use case actually depends on, with ongoing monitoring.

How long does a data readiness assessment take?

For one well-defined use case, usually two to four weeks. The range depends on how many source systems are involved and how quickly owners can answer questions about them. The output should be a ranked list of fixes with effort estimates, not a long report. Assessing the whole organisation at once takes far longer and is rarely needed before the first AI project.

Do we need a data warehouse or lakehouse before starting AI?

Not always. A narrow use case, such as an assistant over a few hundred well-maintained documents, can start without one. Once the use case combines data from several systems, needs history or relies on regular retraining, a governed central store becomes important. Whether that's a warehouse or a lakehouse depends on your data types, your team's skills and your budget.

Can AI tools fix our data quality problems for us?

Partly. AI tools are useful for classifying documents, suggesting likely duplicate records and drafting column descriptions. They can't decide what your business terms mean or who's accountable for them. Use them to speed up the grunt work, have a person review the output, and have the data owner approve any rules that come out of it.

Data & Analytics

Book a data readiness review

Tell us where you're stuck and we'll come back with practical next steps.

Get in touch

Sources