Most AI projects don’t fail because of the tool. They fail because of the data behind it. Here’s what to fix first.

Your AI Tool Isn’t the Problem. Your Data Is.

When a business invests in an AI tool and it underperforms, the instinct is to blame the software. The output is wrong. The recommendations don’t make sense. The automation keeps breaking. So the team either abandons the project or goes shopping for something better.

TL;DR: Gartner predicts that through 2026, organisations will abandon 60 percent of AI projects that are not supported by AI-ready data. The data problems before implementing AI are usually ordinary ones: records scattered across systems, inconsistent formats, stale fields and nobody owning quality. Check those before you sign with any AI vendor.

But in most cases, the tool isn’t the issue. The data feeding it is.

This is the conversation that rarely happens before implementation, which is why so many AI projects quietly die after a promising start.

Gartner put numbers on it in February 2025. In its survey of data management leaders, 63 percent of organisations either did not have, or were unsure whether they had, the right data management practices for AI. Its prediction: through 2026, organisations will abandon 60 percent of AI projects unsupported by AI-ready data. Gartner surveyed 248 data management leaders, so the forecast describes organisations that already employ people to look after data. A small business with no one responsible for data quality starts from further behind.

What “Bad Data” Actually Looks Like

Data quality problems aren’t always obvious. They don’t announce themselves. They show up as AI outputs that are almost right, inconsistently right, or confidently wrong.

Some common patterns:

Scattered data across systems. Customer records in one place, purchase history in another, support tickets in a third. None of them talking to each other. AI models need connected context to make useful inferences. Fragmented data produces fragmented results.

Inconsistent formats and naming conventions. “NSW” vs “New South Wales” vs “New South Wales, Australia.” “Jan 2024” vs “01/01/2024” vs “January 1.” These look like minor annoyances but they create real confusion for models that rely on pattern recognition.

Stale or incomplete records. Outdated contact details, missing fields, records that were never fully populated. An AI tool trained or queried on incomplete data will fill gaps with guesses, and those guesses are often wrong in ways that are hard to catch.

Biased historical data. If your historical data reflects past decisions that were flawed, the AI will learn those flaws. A model trained to recommend products based on previous sales will perpetuate whatever biases existed in those sales, whether that’s regional skews, seasonal anomalies, or gaps in the customer base.

None of these are exotic problems. They’re present in most businesses that have been operating for more than a few years.

Why This Gets Skipped

AI vendors don’t lead with data readiness conversations. The sales pitch is about the output: the automation, the insights, the time saved. Data hygiene is unglamorous, it takes work, and it happens before the technology does anything exciting.

Project timelines don’t budget for it either. Most implementation plans allocate time for setup, configuration, and training. They don’t allocate time for auditing three years of CRM records or standardising how different departments log customer interactions.

The result: businesses go live with AI on top of a shaky data foundation, then wonder why the results don’t match the demo.

What to Do Before You Go Live

A data readiness check doesn’t need to be a multi-month audit. For most SMBs, a focused review of a few key areas will surface the biggest problems:

Map your data sources. List every system that holds data relevant to your AI use case: CRM, accounting, email, spreadsheets, whatever exists. Understand what’s in each and whether they can be connected or exported in a usable format.

Check completeness on your most important fields. If your AI tool needs customer industry, company size, or purchase history to function well, how complete is that data? Even a rough percentage will tell you whether you have a problem worth addressing before launch.

Standardise where it matters most. Full normalisation across every system is a long project. But you can standardise the fields your AI will actually use. Pick those, agree on a format, and clean them before going live.

Set a data ownership rule. Bad data is usually a people and process problem, not a technology one. Someone needs to own each data set and be responsible for its quality. Without that, the problems that existed before the AI will return within months of launch.

A One-Page Data Readiness Check

For a single AI use case, these six questions will surface most of the data problems you need to fix before implementing AI. If you cannot answer one without someone counting records by hand, that is your first job.

CheckQuestion to answerWarning sign
SourcesWhich systems hold the data this use case needs, and can each one export it or connect through an API?Key data lives in a spreadsheet or an inbox with no export
CompletenessWhat share of records have the fields the AI depends on?Nobody knows without counting by hand
ConsistencyIs the same thing recorded the same way in every system?One state, product or client spelt three different ways
FreshnessWhen was each record last updated, and by whom?Contact or pricing data more than a year old still in daily use
OwnershipWho is accountable for the quality of each data set?No named owner, or the answer is everyone
PrivacyCan this data go to the AI vendor, and where will it be stored?Personal information heading offshore with no privacy review

Raise the privacy row with each vendor directly. We list the questions to ask in what to ask an AI vendor before you sign. If the sources row turns up old systems with no API, legacy systems are often the real blocker, and fixing the connections comes before choosing a model.

The Payoff

Getting this right before implementation, rather than retrofitting it after problems emerge, dramatically changes the trajectory of an AI project. Clean, connected, consistent data means your AI tool behaves predictably. It means outputs you can trust. And it means the time your team invests in learning and adopting the technology isn’t wasted on chasing down errors.

The businesses that see real returns from AI aren’t necessarily using better tools. They’ve usually done the unglamorous work of getting their data in order first.

Where to Start

If you are planning an AI implementation and have not done a data review yet, that is the first step. Before you sign anything, before you configure anything, before you schedule training. Our 15-question AI readiness scorecard covers data alongside strategy and skills if you want a quick self-assessment first.

It doesn’t need to be perfect. It needs to be honest. An honest picture of your data situation will tell you how long implementation will realistically take, what needs to be fixed upfront, and whether you’re ready to go.

Avatar Studios works with Australian businesses at exactly this stage, before the tool is chosen, not after it’s failing. If you’re thinking about where AI fits in your business, start with a conversation.

Frequently Asked Questions

What does AI-ready data mean for a small business?

Data that is accurate, complete enough for the use case, consistently formatted, reachable through an export or API, and owned by someone accountable for its quality. It does not need to be perfect across the whole business, only in the fields the AI tool will use.

How long does a data readiness check take?

For a single AI use case, a focused review is a matter of days rather than months: map the sources, measure completeness on the key fields and agree formats. Larger clean-up work, such as merging duplicate CRM records, is scoped separately once you know its size.

Can an AI tool clean up our data for us?

Partly. AI is useful for spotting duplicates, standardising formats and flagging gaps, but a person still has to decide what the correct value is and own the result. Feeding AI-cleaned data back into another AI system without checking it compounds errors.

Does bad data create privacy risk as well as accuracy problems?

Yes. If your business is covered by the Privacy Act, Australian Privacy Principle 10 requires reasonable steps to ensure the personal information you use or disclose is accurate, up to date, complete and relevant to the purpose. Sending stale or duplicated customer records to an AI vendor creates both an accuracy problem and a compliance one.

Related reading

Once your data is in shape, the next questions are strategic: why strategy comes before software, building a team that uses AI well, and the business owner’s guide to AI strategy.