Data quality: the unglamorous work that decides whether AI helps
Every AI project stands on data. Five dimensions of quality, how to measure them in an afternoon, and the fixes that pay off before any model.
The short answer
Every AI project stands on data, and AI applied to poor data does not fail loudly. It produces confident, fluent, wrong output: a summary that misses the amended contract, an answer built on a duplicate record, a forecast from a pipeline nobody updated. The model is rarely the problem. The data underneath usually is. Before any model or vendor is chosen, measure five dimensions of data quality on a sample, fix the unglamorous basics, and put an owner on each dataset. It is the cheapest work in any AI project and the most frequently skipped.
The five dimensions
| Dimension | The question | How to measure on a sample |
|---|---|---|
| Completeness | Are the fields that matter filled in? | Take a hundred records; count empty values per field |
| Accuracy | Do the values match reality? | Check thirty records against the source: the invoice, the contract, the customer |
| Consistency | Is the same fact recorded the same way everywhere? | Look for the same customer or product across systems; count the variants |
| Timeliness | Is the data current enough for the use? | Check when records were last updated versus when they changed in reality |
| Accessibility | Can the data be reached by a system, with permission, in a usable format? | Try to export it; note formats, locks, missing access |
The fixes that pay off
- Required fields at entry. The cheapest fix for completeness: a record cannot be saved without the fields the business depends on.
- One source of truth per fact, with the other systems reading from it rather than holding their own version.
- Deduplication once, then prevention. Merge the duplicates, and add matching at entry so they do not return.
- Definitions written down. What “active customer” means, what the status values are, what a date field records. Half of inconsistency is people meaning different things.
- An owner per dataset, named, who decides definitions, access and entry rules.
- A quarterly sample check on the five dimensions, so decay is caught early.
Matching the effort to the use
Not every AI use needs every dimension. Summarising documents needs accessibility and timeliness: can the system reach the current version? Answering questions from company knowledge needs accuracy and consistency: one correct answer per question, not three. Automating a process needs completeness in the fields the process reads. Forecasting needs all five, over a long history, which is why it is usually the last thing worth attempting. Measure what the intended use depends on, and fix that first.
What this means for you
Before the model, before the vendor, before the pilot: sample the data and measure the five dimensions. Fix entry rules, name owners, choose sources of truth, deduplicate. Then match the AI use to the data that can support it. The projects that fail are almost never short of clever models. They are short of data that means what everyone assumed it meant.
Frequently asked questions
How good does our data need to be before AI is worth trying?
Good enough for the specific use. Summarising documents needs the documents to be findable and current. Answering questions from a knowledge base needs that base to be accurate and deduplicated. Forecasting needs complete history. Measure the dimensions that the use depends on, not all of them everywhere.
Can AI clean the data for us?
It can help with some steps: matching duplicates, extracting fields from text, flagging inconsistencies for a person to check. It cannot decide which of two conflicting records is true or fill in facts nobody recorded. Use it as an assistant to the cleaning, with a person deciding.
Who should own data quality?
Each dataset needs one named owner who decides what the fields mean, who may change them and how new data enters. Without an owner, quality decays back within months of any clean-up. It is a role, not a project.