Why poor data kills AI projects

When AI projects in business disappoint, it’s rarely because the AI isn’t clever enough. It’s because the information underneath is duplicated, inconsistent or incomplete, and the AI does exactly what it was asked with whatever it was given.

The National AI Centre puts it plainly: “An AI system depends on the information you give it. If data is incomplete, outdated or inconsistent, the results will reflect those weaknesses.”

What makes it different from a normal report is how the mistake looks. A bad report often looks bad. Numbers don’t add up, a total is obviously off, and someone notices. An AI answer built on bad data reads well. It’s fluent, confident and plausible, which makes it much harder to catch.

The OAIC’s guidance on AI products makes the same point: “AI outputs can often appear credible even when they produce errors or false information.”

A wrong report gets questioned. A wrong answer that reads well gets forwarded.

How bad data shows up in AI answers

• A customer split across three records looks like three small customers, so the AI misses your biggest account
• Products coded inconsistently get compared as if they were different things
• Old contracts still sitting in a folder are read as if they were current
• Costs allocated differently at each site make one site look far better than it is
• Missing fields get skipped quietly, and the answer is built on whatever was left

None of these produce an error message. They produce an answer that is slightly or completely wrong, delivered in the same confident tone as a correct one.

Why this catches businesses out

Most businesses have learned to work around their data. The analyst knows that customer 1043 and customer 2291 are the same company. The financial controller knows to leave out the old cost centre. That knowledge lives in people’s heads and gets applied without anyone thinking about it.

An AI system doesn’t have that knowledge unless someone gives it. It reads the data as it is.

What to do before an AI project

• Pick the questions the AI needs to answer first, and look only at the data those questions depend on
• Check that data for duplicates, gaps and inconsistent codes
• Write down the workarounds people already use, and turn them into rules
• Remove out-of-date documents, or mark them clearly
• Test the AI on questions you already know the answer to, before anyone relies on it

The last step is one the National AI Centre’s guidance for AI adoption lists among its foundations: test before you deploy a system.

This doesn’t need a major data program. Most AI work in a business depends on a handful of core data sets: customers, products, suppliers, costs and contracts. In our experience, getting those into shape is usually a matter of weeks.

Good AI shows its working

The other protection is an AI system that shows where its answer came from. If it can point to the documents and records it used, a person can check the answer quickly and spot when something is off. An answer with no trail has to be taken on trust, and that’s a poor basis for a business decision.

You don’t have to fix everything first

Some advice says to fix all the data first and start on AI later. In practice that can hold up useful work for years. A better approach is to start AI on a focused problem, find the data issues it exposes, fix those, and widen from there. The AI ends up showing you which data matters most.